CN100550040C - Optical character recognition method and equipment and character recognition method and equipment - Google Patents

Abstract

The invention relates to an optical character recognition method and equipment. In one aspect, the present invention provides an optical font recognition method and device as follows: the words of the input text image are divided into word pairs, the longer words and the shorter words in the corresponding word pairs are respectively recognized, respectively, according to the adjacent The font information of the word and the coarse adjustment step of adjusting the font information of the word according to the distribution of the font information in the row adjust the font information of the word, and additionally identify the font size of the word in the row. The present invention also provides a font recognition method and device based on a classification process and a method and device for identifying interval types and calculating X-height by using a projection method combined with a connected unit method.

Description

Optical character recognition method and device, font recognition method and device

technical field

The present invention relates to an optical character recognition method and equipment, more specifically, to a method and equipment for identifying character font types in an optical character recognition system.

Background technique

Optical character recognition (OCR) systems have been widely used. Font information such as font style, slant, point weight and font size have been used in conventional OCR systems to improve the performance of OCR, and font information is also beneficial to the performance of file structure analysis and information recovery.

There are two methods available for font recognition today:

-Extract global features from textual entities (words, lines, paragraphs). This approach is suitable for a priori font recognition, where fonts are identified without any knowledge of letter classes.

- Extract local features from individual letters. This approach can benefit substantially from knowledge of letter classes.

In particular, US6,337,924 and US6,496,600 disclose a posteriori font recognition. First, the character category (code) is known, from which local features can be extracted to determine the font of the specified character. These features are based on letter details such as the shape of the serif (arch, square, triangle, etc.) . However, it is a completely different type of font recognition that can use knowledge of letter classes. This type of approach can benefit substantially from knowledge of letter classes and thus achieve higher accuracy. But it is not helpful for predicting font information for full-font OCR.

CN 1271140A (Chinese application number 99105851.8) and "Optical FontRecognition Using typographical features" (IEEE TRANSACTIONSON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL.20, NO.8, AUGUST 1998) to Abdelwahab Zramdini gold and Rolf have disclosed a block based on text Or line "English Font Recognition". Global features are extracted to determine font information for an entire text block or line. These characteristics are generally determined by non-printers (such as text density, size, orientation and spacing of letters, serifs, etc.). However, different words in a line can be in different fonts (typeface, size, point, slant, etc.), such as names and weights, etc. This method generally extracts features from a large block of text (such as a text block, at least one line), and cannot recognize the font information of a single word.

Printed in India)已经公开了一种基于词的“英文字体识别”。字体特征从每个词图像中提取，因此它可以识别文本行中每个词的不同字体信息。然而，其它英文单词OFR处理通常不利用字形识别中的X高度或区间类型信息。但是在理想的英文单词OFR中，词X高度是非常有用的信息。词图像可以根据X高度归一化并且从归一化的图像中提取字体特征，以便确保以相同的等级提取不同词的字体特征并避免字号的影响。词区间类型也是字形识别所需的信息。通常，不同区间的英文单词类型可以具有不同的字体特征，这是因为在不同的区间中不同字符(描边)组合的缘故。词X高度和区间类型也用于字号识别。一般其它字体识别处理还没有确保精确的X高度和词区间类型计算的成熟的方案。它们通常通过简单的投影直方图法获得行的信息。Optical font recognition from projection profiles (Electronic publishing, VOL.6(3), 249-260 (September 1993)) and Script identification in printed bilingual documents to D DHANYA, A GRAMAKRISHNAN_ and PEETA BASA PATI (Sadhana Vol.27, Part1, February 200 , pp.73-82.

Printed in India) has disclosed a word-based "English font recognition". Font features are extracted from each word image, so it can identify different font information for each word in a text line. However, other English word OFR processes generally do not utilize X-height or interval type information in glyph recognition. But in the ideal English word OFR, the word X height is very useful information. Word images can be normalized according to the X height and font features are extracted from the normalized image, so as to ensure that font features of different words are extracted at the same level and avoid the influence of font size. The word interval type is also information required for font recognition. Generally, English word types in different intervals may have different font characteristics, which is due to different combinations of characters (strokes) in different intervals. Word X height and interval type are also used for font size recognition. In general other font recognition processes do not yet have a mature solution for ensuring accurate X-height and word space type calculations. They usually obtain row information by simple projection histogram method.

In addition, fonts may be changed at specific points in the document for emphasis or to ease the reader's attention. This can be done by selecting another font or changing the style (such as point size, slope, etc.) or choosing a different font size of the same font. In some of the previous font recognition methods described above, font features are extracted from blocks or lines of text, or from individual words. Obviously, for a text line or block of a single font, the first method is easy to obtain font information, but cannot identify the font information of a single word, and different words in a line can have different fonts (font style, font size, point number, slope, etc.). The second method can identify different font information for each word in a text line, but it cannot exploit contextual font information, and usually they are not completely irrelevant. Before performing any OFR processing, special consideration must be given to the size of the region to be processed. If the region is too small, there may not be enough information contained in it for classification; however, if it is too large, there may be too many different types mixed in the same region.

Contents of the invention

In view of the above-mentioned defects of the prior art, it is required to provide a novel font recognition method capable of eliminating the above-mentioned defects of the prior art, and this method can accurately identify the font information (font shape, font size, serif, point size, slope, and spacing).

Also require a novel font recognition method that can recognize more fonts than common OCR software (such as Omnipage, FineReader, etc.) with file layout recovery functions.

According to a first aspect, the present invention provides an optical font recognition method, comprising the steps of:

The step of dividing the words of the input text line image into word pairs;

A font recognition step that identifies font information of the longer word in each word pair;

Discriminating the font information of the shorter word in each word pair based on the font information of the longer word in the word pair that includes the shorter word and the font information of the longer word in the word pair adjacent to the shorter word. Font identification step.

Preferably, a fine adjustment step of adjusting the font information of words according to the font information of adjacent words is also included.

The optical font recognition method further includes a recognition step of recognizing font sizes of words in the above-mentioned lines.

This optical font recognition method uses a word pair mechanism in the OFR processing of line images, which is based on the font category method of English words and simultaneously considers the characteristics of font distribution in actual English texts. It uses a two-stage result conditioning technique based on contextual font information, which enables greater precision and neater output within lines of text.

Preferably, the font identification step of identifying the font information of the shorter word in each word pair comprises:

comparing the font information of the longer word in the word pair adjacent to the shorter word to be identified with the font information of the longer word in the word pair in which the shorter word is to be identified; and

If the font information of the longer word in the word pair adjacent to the shorter word to be identified is the same as the font information of the longer word in the word pair where the shorter word is located, determine the shorter word in the word pair have the same font information as the longer word in the word pair and the longer word in the adjacent word pair.

Preferably, if the font information of the longer word in the word pair adjacent to the shorter word to be identified is different from the font information of the longer word in the word pair where the shorter word to be identified is located, then identify the font of the shorter word information.

Preferably, the step of identifying the font information of the shorter word in each word pair further comprises identifying the font information of the shorter word if the shorter word in the word pair is the first or last word in the line.

Preferably, the fine-tuning step of adjusting the font information comprises: determining whether a first candidate font for each word in the line is available; if it is determined that the first candidate font for the word is not available, determining whether a second candidate font for the word is available ; If it is determined that the second candidate font of the word is available, the second candidate font of the word is exchanged with the first candidate font of the word.

Preferably, the step of fine-tuning the font information further includes determining whether a third candidate font for the word is available if it is determined that the second candidate font for the word is not available, and determining whether the third candidate font for the word is available if it is determined that the word is available. The third candidate font of is exchanged with the first candidate font of the word.

Preferably, if it is determined that the third candidate font of the word is unavailable, then it is judged whether all three candidate fonts of the word are reliable; and if it is determined that all three candidate fonts of the word are unreliable, the corresponding The font of the adjacent word sets the first candidate font for the word.

Preferably, the fine-tuning step includes the following preprocessing process: compare the font information of each word in the row with a dictionary containing font information of a plurality of known fonts; and obtain three candidate fonts and three candidate fonts of the word in order of similarity A corresponding distance value, the condition for judging that all three candidate fonts of the word are unreliable includes that all three distance values are greater than a predetermined threshold.

Preferably, the fine-tuning step includes the following preprocessing: comparing the font information of each word in the row with a dictionary containing font information of a plurality of well-known fonts; candidate fonts and at least two corresponding distance values; and counting the distribution of detailed glyphs in the line according to the first candidate font for the word.

Preferably, for the first or last word in a line, the conditions for determining whether the first candidate font of the word is available include: the detailed glyphs are roughly consistent in the line, or the first candidate font of the word is the same as the main font of the line The font is consistent.

Preferably, for the first or last word in the row, the conditions for determining the availability of the second candidate font of the word include: the second candidate font of the word is consistent with the main glyph of the row and the corresponding distance value is less than a predetermined threshold .

Preferably, the candidate fonts include a first candidate font, a second candidate font and a third candidate font, and for the first or last word in a row, the conditions for determining the availability of the third candidate font for the word include: word The third candidate font is consistent with the main glyph of the row and the corresponding distance value is smaller than a predetermined threshold.

Preferably, for each word except the first and last words in the line, the conditions for determining whether the first candidate font of the word is available include: the detailed fonts are roughly consistent in the line, and the first word of the current word A candidate font is consistent with the main glyph of the row, or the first candidate font of the current word is different from the first candidate fonts of two adjacent words in the row, or the detailed glyphs of the two adjacent words are the same.

Preferably, for each word except the first and last words in the line, the conditions for determining that the second candidate font of the word is available include: the second candidate font of the word is consistent with two adjacent words and The corresponding distance value is less than a predetermined threshold.

Preferably, the candidate fonts include a first candidate font, a second candidate font and a third candidate font, and for each word except the first or last word in a line, it is determined that the third candidate font for said word is available The conditions include: the third candidate font of the word is consistent with the main glyphs of two adjacent words and the corresponding distance value is smaller than a predetermined threshold.

Preferably, it also includes a coarse adjustment step of adjusting the font information of the row according to the distribution of font information in the row, and the coarse adjustment step of adjusting the font information of the row includes: at least one of serifs and spacing in the row and detailed glyphs distribution; determine whether said at least one of serifs and spacing is uniform and whether detail glyphs are approximately consistent in a row, and if it is determined that the above-mentioned distribution satisfies this condition, set all word's first candidate font and use each word's first candidate font as the identified font result.

Preferably, the recognition step of identifying the character size of the word in the row includes: judging whether the input word X height of the word is available; if the input word X height is available, querying with the recognized font and input image resolution of the known word contains " Image resolution/font style/font size/X height" table's prior font size dictionary, and get a list of X heights of different font sizes; match the X height of the word with the X height in the X height list; and set the corresponding font size as the recognized font size.

Preferably, the recognition step of identifying the font size of the word in the line includes: querying a prior font size dictionary containing the table "image resolution/font shape/font size/word height" with the recognized font shape of the known word and the input image resolution, and A list of word heights of different font sizes is obtained; the word height is matched with word heights in the word height list; and the corresponding font size is used as the recognized font size.

According to a second aspect, the present invention provides a font recognition method, comprising:

A normalization step that normalizes the words of the input image to a predetermined height;

A feature extraction step that extracts features from normalized words;

The step of judging the interval type of the judgment word;

classification steps, and

A recognition step of recognizing font information of words based on the result of the classification step.

This font recognition method normalizes the height of the word image to a given height, thereby extracting the features of the word image with the same interval type at the same level. sort. First consider the word interval type to select the corresponding interval type dictionary in the grapheme classification. When the resulting distance is greater than a given threshold or the word interval type is unknown, a priori permuted multi-dictionary recognition process will be used. Therefore, such an embodiment will take full advantage of the information about the interval type, and the overall OCR process becomes more efficient.

Preferably, the input image is a word image from a document image that has undergone line and word segmentation processing.

Preferably, the judging step includes a step of judging whether the interval type of the word is known from the outside.

Preferably, the classification step further includes comparing the extracted features of the normalized word with the features of the candidate glyphs in the judged word interval type dictionary, and obtaining a distance value from the comparison, and the method further includes before the recognition step A verification step of verifying the distance value with a predetermined threshold. Bayesian classifiers are preferably used here for the classification step.

Preferably, the dictionary is selected from at least a dictionary of X-height words, a dictionary of all-caps words, a dictionary of ascenders and descenders, and a dictionary of full-height words.

Preferably, each dictionary of interval type has at least 40 types of candidate glyphs.

Preferably, if the interval type of the word cannot be obtained in the adjusting step or if the distance value is smaller than a predetermined threshold in the verifying step, the method further comprises a subdivided classification step using a plurality of dictionaries.

Preferably, the subdivision step includes: comparing the extracted features of the normalized word with the features of the candidate glyphs in the dictionary of at least one other word interval type except the judged word interval type, respectively from at least one comparison step A distance value is obtained and verified against another predetermined threshold to identify the font information of the word based on the result of the verifying step.

Preferably, the at least one word interval type dictionary to be compared includes dictionaries in the following order: a dictionary of X-height words, a dictionary of all-caps words, a dictionary of ascenders and descenders, and a dictionary of full-height words.

Preferably, the font information of the word to be recognized includes at least font style, serif, point size, slope and spacing.

Alternatively, the present invention also provides another method for font recognition, including:

A normalization step that normalizes the words of the input image to a predetermined height;

A feature extraction step that extracts features from normalized words;

a refinement step using multiple dictionaries; and

A recognition step of recognizing font information of words based on the result of the subdivision step.

Similarly, the input images are word images from document images that have been processed by line and word segmentation, and Gabor filters are used to perform the feature extraction step and Bayesian classifiers are used for the fine-grained classification step. In addition, the fine classification step includes: comparing the extracted features of the normalized word with the features of the candidate glyphs in the dictionary of at least one word interval type, respectively obtaining distance values from at least one comparison step, and verifying the distance with a predetermined threshold value to identify the word's font information based on the result of the verification step. The dictionaries of at least one word interval type to be compared include dictionaries in the following order: a dictionary of X-height words, a dictionary of all-caps words, a dictionary of ascenders and descenders, and a dictionary of full-height words.

Preferably, the font information of the word to be recognized includes at least font style, serif, point size, slope and spacing.

According to a third aspect, the present invention provides a method for identifying an interval type of an input text line, comprising:

A line information calculation step for calculating line information of an input text line using a projection method;

A word information calculation step for calculating word information of selected words in the text line using a projection method;

A reliability judgment step for judging whether the word information is reliable;

The interval type marking step of marking the interval type of the word according to the line information and the word information;

Wherein, if it is judged in the reliable judgment step that the calculated word information is not reliable, then the connected unit method is used to calculate the word baseline and upper line of the selected word.

This method of identifying the range type of an input text line combines two approaches to improve the accuracy of the X-height value calculation. Because the projection method is much faster than the connected unit method, the projection method is firstly used to calculate the word on-line, and the connected unit method is used when the reliable line information of the word cannot be obtained by the projection method. Perform rigorous checks and checks several times to get more and more accurate counted words on the line. Finally, the reliable word X height value and interval type are obtained.

Preferably, the line information calculation step includes: performing vertical projection in the entire word area to obtain the baseline and upper line of the text line. And the row information calculating step further includes: checking the reliability of the obtained row baseline. Checking the reliability of the obtained row baseline includes: if the height of the lower interval located between the row baseline obtained from the row information calculation step and the row bottom line is not less than the height between the row upper line obtained from the row information calculation step and the row baseline The height of the middle interval of , then perform the vertical projection method again in the interval below the rough baseline to obtain a new row baseline.

Preferably, the word information calculation step includes: using a vertical projection histogram to calculate the word upper line in the upper half of the word interval; and the step of calculating the word baseline based on the relationship between the height of the word and the y coordinate of the row baseline, and if A row baseline close to the bottom of the word directly determines that the word baseline is equal to the bottom of the word.

Preferably, the reliability judging step judges whether the word line is reliable by determining the relationship between the line information and the word information.

Preferably, the section type marking step includes preliminary marking of the selected word's section type using classification conditions based on line information and word information. The classification criteria include selecting at least one feature from the following group:

z1/(z1+z2);z3/(z3+z2);nzw;th;|ul_w-ul_l|/(z1+z2);

Here, z1 refers to the height of the upper interval of the word, and z2 refers to the height of the interval in the middle of the word; z3 refers to the height of the lower interval of the word; ul_w refers to the y coordinate of the upper line of the word (in the word coordinate system); ul_l refers to the y coordinate of the upper line of the line ( In the word coordinate system); nzw refers to the ratio of w1 to w2; w1 refers to the width of the non-zero interval of the horizontal projection histogram in the upper interval of the word; w2 refers to the width of the word; th refers to the ratio of max1 to max2; max1 refers to The maximum value of Pv [i] {i so that it is close to the top line of the word}; max2 refers to the maximum value of Pv [i] {i so that it is close to the top line of the word}. Preferably, the feature nzw is used in combination with another feature th.

Interval types of words selected by markers may include: X-height words; ascenders or whole words; descenders or whole words; words above X-height whose upper line is close to the line's upper line; words above X-height; and unknown word.

Preferably, the method for identifying the interval type of the input text line may further include the step of judging whether the selected word in the text line is a short word and calculating the word information of the judged short word by using a projection method in a designated area of the word interval . Therefore, the method of using the interval type of the input text line can further include the step of verifying that the word information calculated in the word information calculation step is online, and if the word information calculated in the reliability judgment step is reliable, then after verifying the above-mentioned word online After the step, it is judged whether the word in the text line is a short word.

Preferably, the connected unit method comprises at least one of the following steps:

Because these connected units are regarded as noise, delete connected units smaller than a predetermined smaller area;

Delete connected units below the word midline; and

Connected units above the word midline obtained in the word information calculation step are deleted.

Preferably, the section type marking step further includes a step of judging whether to correct word line information according to the marked section type.

Based on the interval type obtained by the above method, it is easy to calculate the X-height of any word in the input text line.

In addition, the present invention further provides an optical font recognition device, comprising:

dividing the words of the input text line image into word pairs;

font recognition means for recognizing the font information of the longer word in each word pair;

Discriminating the font information of the shorter word in each word pair based on the font information of the longer word in the word pair that includes the shorter word and the font information of the longer word in the word pair adjacent to the shorter word. font identification device;

Fine adjustment means for adjusting font information of words according to font information of adjacent words; and

coarse adjustment means for adjusting the font information of the row according to the font information distribution in the row; and

A recognition device that recognizes the font size of the words in the line.

Correspondingly, the present invention further provides a font recognition device, comprising:

normalization means for normalizing the words of the input image to a predetermined height;

A feature extraction device for extracting features from the normalized words;

A judging device for judging the interval type of a word;

sorting device, and

Recognition means for recognizing font information of words based on the results obtained by the classifying means.

Alternatively, the above font recognition device may be modified to include: normalization means for normalizing the input image to a predetermined height; feature extraction means for extracting features from the normalized characters; direct use of multiple dictionaries fine classifying means; and recognizing means for recognizing font information of words based on the result obtained by the fine classifying means.

Therefore, the present invention further provides a device for identifying the interval type of an input text line, comprising:

A line information calculation device for calculating line information of an input text line using a projection method;

a first word information calculating means for calculating word information of words selected in the text line using a projection method;

A reliability judging device for judging whether word information is reliable;

second word information calculating means for calculating word baselines and upper lines of selected words not judged as reliable by the reliability judging means by using a connected unit method; and

A section type marking device for marking the section type of the word according to the line information and the word information;

Therefore, the present invention further provides a device for calculating the X-height of a word in an input text line from the device for identifying the range type of the input text line using the interval type obtained by the device for identifying the interval type of the input text line.

The present invention also provides a program for performing the steps of optical font recognition, the steps comprising:

The step of dividing the words of the input text line image into word pairs;

A font recognition step that identifies font information of the longer word in each word pair;

Discriminating the font information of the shorter word in each word pair based on the font information of the longer word in the word pair that includes the shorter word and the font information of the longer word in the word pair adjacent to the shorter word. font identification step;

A fine adjustment step of adjusting the font information of a word according to the font information of adjacent words; and

a coarse adjustment step of adjusting the font information of the row according to the font information distribution in the row; and

A recognition step that recognizes the font size of the words in the line.

The present invention also provides a program for performing the steps of font recognition, the steps comprising:

A normalization step that normalizes the words of the input image to a predetermined height;

A feature extraction step that extracts features from normalized words;

The step of judging the interval type of the judgment word;

classification steps, and

A recognition step of recognizing font information of words based on the result obtained by the classification step.

Alternatively, the present invention also provides a program for performing the steps of font recognition, the steps comprising:

a normalization step that normalizes the input image to a predetermined height;

A feature extraction step that extracts features from the normalized word;

a subdivision step of directly using a plurality of dictionaries; and a recognition step of recognizing font information of a word based on a result obtained through the subdivision step.

The present invention also provides a program for performing the steps of identifying interval types in input text lines, the steps comprising:

A line information calculation step for calculating line information of an input text line using a projection method;

A first word information calculation step of calculating the word information of the selected word in the text row using a projection method;

A reliability judgment step for judging whether the word information is reliable;

The interval type marking step of marking the interval type of the word according to the line information and the word information; and

Wherein if the calculated word information is judged to be unreliable in the reliable judging step, the word baseline and upper line of the selected word are calculated using the connected unit method.

The present invention also provides a program for performing the steps of calculating the X-height of a word in a line of input text, said steps comprising:

A line information calculation step for calculating line information of an input text line using a projection method;

A first word information calculation step of calculating the word information of the selected word in the text row using a projection method;

A reliability judgment step for judging whether the word information is reliable;

If in the reliability judgment step, it is judged that the word information calculated is unreliable, the word baseline and the upper line of the selected word are calculated using connected units;

The interval type marking step of marking the interval type of the word according to the line information and the word information; and

Calculate the X-height of the word using the obtained interval type.

In another aspect, the present invention also provides a storage medium storing a program for performing the following steps of optical font recognition, including:

The step of dividing the words of the input text line image into word pairs;

A font recognition step that identifies font information of the longer word in each word pair;

Discriminating the font information of the shorter word in each word pair based on the font information of the longer word in the word pair that includes the shorter word and the font information of the longer word in the word pair adjacent to the shorter word. font identification step;

A fine adjustment step of adjusting the font information of a word according to the font information of adjacent words; and

A coarse adjustment step that adjusts the font information of a row according to the distribution of font information in the row.

In another aspect, the present invention also provides a storage medium storing a program for executing the following steps of the interval type of input text line, including:

A line information calculation step for calculating line information of an input text line using a projection method;

A first word information calculation step of calculating the word information of the selected word in the text row using a projection method;

A reliability judgment step for judging whether the word information is reliable;

The interval type marking step of marking the interval type of the word according to the line information and the word information; and

Wherein if the calculated word information is judged to be unreliable in the reliable judging step, the word baseline and upper line of the selected word are calculated using the connected unit method.

In another aspect, the present invention also provides a storage medium storing a program for performing the steps of calculating the X-height of a word in a line of input text, the steps comprising:

A line information calculation step for calculating line information of an input text line using a projection method;

A first word information calculation step of calculating the word information of the selected word in the text row using a projection method;

A reliability judgment step for judging whether the word information is reliable;

If in the reliability judgment step, it is judged that the word information calculated is unreliable, the word baseline and the upper line of the selected word are calculated using connected units;

The interval type marking step of marking the interval type of the word according to the line information and the word information; and

Calculate the X-height of the word using the obtained interval type.

Other features and advantages of the present invention will become apparent from the following description in conjunction with the accompanying drawings.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention.

Accompanying drawing 1 shows the overall flow chart of X-height value calculation and interval type identification according to the first embodiment.

Figure 2 shows the details of the vertical projection method for the calculation of the X-height value.

Figures 3 and 4 show the line information results of the projection method, and two features nzw and th are used in the detailed word interval type marking process.

Figure 5 shows the upper line of a short word "The" using "normal projection method" and "projection in the short word designation area" respectively.

Accompanying drawing 6 shows the character line containing the short word "He" belonging to the short word type 2.

Figure 7 shows the short word "He" processed using a vertical projection histogram.

FIG. 8 is an expanded flowchart of step 600 in FIG. 1 .

Figure 9 shows an application of the identified interval type involved in the first embodiment of the present invention.

Figure 10 shows an application of the calculated X-height value of English words obtained according to the first embodiment of the present invention.

Accompanying drawing 11 shows the main flowchart of the second embodiment of the present invention.

FIG. 12 is an extended flowchart of step 400 in FIG. 11 .

FIG. 13 is an extended flowchart of step 600 in FIG. 11 .

Accompanying drawing 14 shows the flowchart of the application of the second embodiment of the present invention.

Accompanying drawing 15 shows the main flowchart of the third embodiment of the present invention.

Fig. 16 is a detailed main flowchart of step 400 of Fig. 15 according to the third embodiment of the present invention.

Fig. 17 is a detailed flowchart of step 500 of Fig. 15 according to the third embodiment of the present invention.

FIG. 18 is a detailed flowchart of step 600 of FIG. 15 according to the third embodiment of the present invention.

FIG. 19 is a detailed flowchart of step 700 of FIG. 15 according to the third embodiment of the present invention.

Accompanying drawing 20 shows the flowchart of the application of the third embodiment of the present invention.

Fig. 21 is a flow chart of OCR according to the fourth embodiment of the present invention.

Fig. 22 shows the device for implementing the method for optical character recognition according to the third embodiment of the present invention.

Fig. 23 shows the device for implementing the font recognition method according to the second embodiment of the present invention.

Fig. 24 shows the device for implementing the font recognition method according to the second embodiment of the present invention.

Fig. 25 is a diagram showing the apparatus for the method of identifying the interval type of an input text line according to the first embodiment of the present invention.

Detailed ways

The optical character recognition method and apparatus according to the present invention are described below by referring to the accompanying drawings.

Terms appearing in the present invention are explained as follows:

· OFR: Optical Font Recognition.

OCR: Optical Character Recognition

• ICR: Individual character recognition (OCR of individual characters).

• X-height: the height is the portion of a letter of the same height as a lowercase letter x of the same glyph.

·Ascending letters: The part of the lowercase letter that rises above the X height is the ascending part.

Descenders: The part of a lowercase letter that protrudes below the baseline is the descender.

Glyph: A variation of a style within a font family type, such as Roman, Italic, Bold, Ultra, Simplified, Expanded, Shaped, Outlined, etc.

Point size: The point size of a character is determined by the relationship between the thickness of its stroke (stroke) and its overall height. Most glyphs are set to two weights, normal and bold.

• Slope: The slope indicates the orientation of the main stroke of the letter. The font can be roman or italic.

• Serifs: the short cross-hatches at the ends of the ascenders and descenders of certain glyphs.

· Spacing: Spacing refers to the horizontal space required by each character of a glyph. Because different characters have different widths, different spacing adjusts the gap between characters. Single spacing gives each character the same space regardless of width.

[Example 1]

Embodiment 1 accurately calculates the X-height value of an English word, and reliably identifies the interval type (ascending letter, descending letter, full height or X-height) of an English word.

This embodiment improves the projection method and connected cell (CC) method, and combines them to improve the accuracy of X-height value calculation.

It is well known that vertical projection distributions as disclosed in Optical font recognition from projection profiles (Electronic publishing, VOL.6(3), 249-260 (September 1993)) can only handle words or lines of text with ideal vertical projection distributions , and its performance deteriorates when dealing with short words and distinguishing between X-height and uppercase characters. This method may fail when the text lines are skewed. In addition, it is difficult to identify the interval types of English words only by the vertical projection distribution, because sometimes the projection distributions of words with different interval types are similar in shape and thus difficult to classify.

However, because the projection method is much faster than the connected unit method, the projection method is used to quickly calculate the word on-line, and the connected unit method is used when the reliable line information of the word cannot be obtained by the projection method. Rigorous checks and verifications are performed multiple times to make the counted words on the line more and more accurate. Finally, reliable word X-height values and interval types are obtained.

Accompanying drawing 1 shows the overall flow chart of X-height value calculation and interval type identification according to the first embodiment.

In step 100, the row baseline is flexibly calculated using a projection method to avoid skewing. Accompanying drawing 2 shows the details of the vertical projection method used for the calculation of the X height value, the height of the top line (Htop), the height of the bottom line (Hbottom), the height of the baseline (Hbase) and the height of the upper line (Hupper) calculation as follows:

Htop=max{is to make Pv[i]>0};

Hbottom=min{i so that Pv[i]>0};

Hbase=i to maximize Pv[i]-Pv[i-1];

Hupper=i to minimize Pv[i]-Pv[i-1].

First, a vertical projection is performed across the text area to obtain a rough baseline and upper line of the text line.

Then, check the validity of the row baseline. If the height of the lower interval between the row baseline and the row bottom line is smaller than the height of the middle interval between the row upper line and the row baseline, the row information may be correct. Otherwise, perform the vertical projection method again in the interval below the coarse baseline to obtain a more reliable row baseline in order to avoid the effect of row skew.

In step 200, the vertically projected histogram is used to calculate the upper line of words in the upper half of the word interval.

In step 300, this embodiment calculates word baselines using different methods based on different row baseline results. Details are shown in Table I.

Table I

No. Row baseline results Corresponding method 1 0.57＊Hw-1<Y(baseline)<Hw-2, (this baseline is in the lower half of the word interval) Vertically project the histogram in word intervals around (±20%) the row baseline 2 |Y(baseline)-(Hw-1)|≤1, (the baseline of this line is close to the bottom of the word) word baseline equals word bottom 3 other situations Vertically project the histogram in the word interval whose height ranges from 55% to 100% of the word height

here,

Hw: word height

Y (baseline): the y-coordinate of the row's baseline

If 0.57*Hw-1 < Y(baseline) < Hw-2, which means the row baseline is in the lower half of the word interval, then use the vertical projection histogram in the word interval around the row baseline (within ±20% of the baseline) graph to calculate word baselines; and

If Y(baseline)-(Hw-1)|≤1, which means that the row baseline is close to the bottom of the word, determine that the word baseline is equal to the bottom of the word; otherwise,

Vertically projected histograms were used to calculate word baselines in word intervals whose heights ranged from 55% to 100% of word height.

In step 350, it is judged that the line information is definitely wrong, and the judgment criteria used are defined as follows:

When one of the following three criteria is not met, the information in the certain line is wrong:

The word on-line is not very close to the top of the line;

The height of the line interval below the word baseline does not exceed the height of the line interval between the upper line line and the word baseline;

• The word baseline is in the interval between 57% and 100% of the word height.

If it is determined in step 350 that the line information is incorrect, go to step 600 and use the Connected Cell (CC) method to calculate the word baseline and upperline (will be described in detail below). If it is determined at step 350 that the line information is correct, then step 380 is used to mark the detailed section type according to the line and word information described above. Details are shown in Table II.

Table II and other places in this specification use the following 6 kinds of interval types: interval type No.1 refers to the interval type of X height words; interval type No.2 refers to the interval type of ascending letters or whole words (full words); interval type No.3 refers to the interval type of descending letters or full words; interval type No.4 refers to the interval type of words whose upper line is close to the upper line of the line and is higher than the X height; interval type No.5 refers to the interval of words higher than the X height type; and interval type No. 6 refers to an unknown word.

Table II:

Interval type serial number. Classification criteria Detailed interval type 1 z1/(z1+z2)＜0.065 and z3/(z3+z2)＜0.065 X height word 2 0.22＜z1/(z1+z2)＜0.42 and nzw＜0.6 or th＜0.6 Ascending letters or whole words 3 0.21<z3/(z2+z3)<0.45 and z1/(z1+z2)<0.42 Descenders or whole words 4  ＜z1/(z1+z2)≤0.22) and (|ul_w-ul_l|/(z1+z2)＜0.065) and (nzw＜0.6 or th＜0.6) Interval type of words higher than X-height whose upper line is close to the upper line of the row 5  ＜z1/(z1+z2)≤0.22) and (nzw＜0.6 or th＜0.6) Words above X-height 6 Other situations unknown word

here,

z1: refers to the height of the upper section of the word between the line top line and the line top line;

z2: refers to the height of the interval in the middle of the word;

z3: refers to the height of the lower part of the word;

ul_w: refers to the y-coordinate of the upper line of the word (in the word coordinate system);

ul_l: refers to the y coordinate of the upper line (in the word coordinate system);

nzw: refers to the ratio of w1 to w2

w1: refers to the width of the non-zero interval of the horizontal projection histogram in the upper interval of the word;

w2: refers to the width of the word;

th: refers to the ratio of max1 to max2;

max1: refers to the maximum value of Pv [i] {i so as to be close to the top line of the word};

max2: Refers to the maximum value of Pv[i] {i so as to be close to the word line}

Here, this embodiment uses two features nzw and th in marking the detailed word interval type. These two features are specified for checking the row information results of the projection method, but are invalid when checking the results of the connected unit method. Detailed interval types can be obtained satisfactorily when the features nzw and feature th are respectively used as features in the case of classification.

Figure 3 shows the line information results of the projection method. The upper hairline in each of the words "were" and "labor" is counted as the word top line, and the lower hairline across each of the words "were" and "labor" is the word top line. Obviously, the word "were" is wrong, but the word "labor" is correct. The following explains how the two features nzw and th work.

In Figure 4, the horizontally projected histogram is applied in the upper interval of the word in order to arrive at "w1", regardless of whether the word is "were" or "labor". Then, get "nzw" easily. The "nzw" value of the word "were" is 0.73, while the value of other words is 0.16. These results are quite different. Therefore, in this embodiment, a threshold is set to determine whether the line information result of the projection method is correct.

On the right side of Fig. 3, vertically projected histograms are applied to obtain the maxima of the histograms near the top line of words and near the top line of words, literally. Then, it is determined that the "th" of the word "were" is 0.90, and the other values are 0.29. These results are also quite different. Therefore, it is easy to set a threshold to judge whether the result is correct or not.

In short, in order to avoid the judgment of a single feature being too strict and prone to errors, the two features are combined for judgment. If both features "nzw" and "th" are larger than the threshold, the row information result of the projection method must be wrong, otherwise it is reliable.

In step 400 as shown in FIG. 1 , the judging criteria used here are as shown in Table III to verify the additional conditions for online words.

Table III

Interval type Additional conditions deal with 1 N/A unverified 2 R＞3.9 or (R＞=1.3 and (z1+z2)/(zz1+zz2)＞0.63)   Verify online 3 N/A   Verify online 4 R＞3.9   Verify online 6 other situations Do not verify

here,

R: aspect ratio = word width/(z1+z2)

zz1: the height of the upper section of the line

zz2: the height of the interval in the middle of the row

If it is determined that the words can be verified in step 400, then go to step 800 and determine whether the last information is correct. However, if it is determined that the word cannot be verified in step 400, then go to step 450 and use the following criteria to determine whether the word is a "short word".

Table IV

here,

max1: refers to the maximum value of Pv[i] {i for the middle interval and the upper third of the upper interval};

max2: Refers to the maximum value of Pv[i] {i so that the middle third of the middle interval and the upper interval}

ul_l: the y coordinate of the upper line (in the word coordinate system);

R: aspect ratio = word width/(z1+z2)

zz1: the height of the upper section of the row

zz2: the height of the interval in the middle of the row

nzw and th: the same judgment criteria used in step 380 .

A word is a "short word" if the above additional conditions are satisfied.

If it is determined that the word is not a "short word", go to step 600 and use the Connected Cell (CC) method to calculate the word baseline and upperline (described in detail below). If it is determined that the word is a "short word", then step 500 calculates the word on-line of the short word with projection method in the designated area (such as the area on the right of accompanying drawing 5), then in step 700 according to the word on-line again to the detailed interval type to mark.

Table V

phrase type Approach 1 Vertically projected histograms in word intervals ranging from 22% to 42% of the upper and middle intervals 2 Vertically project the histogram in word intervals near the line above the line

Figure 5 shows the upper line for the word "The". The upper thin line on the word is the result of the normal projection method, while the lower thin line across the word is the result of "projection in the short word designation area". Therefore, it is easy to judge that the former is wrong and the latter is correct.

As shown in Figure 5, max1 is smaller than max2, so the short word "The" belongs to short word type 1. Then use the corresponding processing method according to Table V: vertical projection histogram in the white specific word interval around the lower thin line across the word. Then, the correct word can be obtained on the line.

As shown in Figure 7, the top line of the short word "He" is close to the top line of the line, and the line information of the entire line meets the specified conditions, so the word belongs to short word type 2.

Then use the corresponding processing method: vertically project the histogram in the word interval near the line on the row as shown by the white specific interval projected from the upper periphery of Fig. 7 . Finally, the correct word overline is obtained, as shown by the upper thin line directly on top of the letter "e" shown in Figure 7, while the wrong word overline is obtained if the normal projection method is used.

Below explain step 600, wherein use connected unit (CC) method to calculate word baseline and upper line, the condition of this step execution is as follows: if in step 350 determine row information is wrong, or if determine word is not " short word " in step 450, That is, when reliable row information of words cannot be obtained by the projection method. Connected cell (CC) methods have been widely used in the art. Here, this embodiment provides a new connected unit method, in which connected units with inappropriate areas or positions are regarded as noise or punctuation, and then eliminated to obtain more accurate word baselines and upper lines.

FIG. 8 shows an expanded flowchart of step 600 in FIG. 1 .

In step 610 of FIG. 8 , the connected unit (CC) whose area is smaller than SmallArea is deleted after the word connected unit (CC) information is marked. Here, SmallArea is chosen to be equal to 0.015 of the word area, because connected units whose area is smaller than SmallArea are considered noise and should be removed. The ratio of SmallArea to word area is selected according to experience or actual requirements.

In step 620 of FIG. 8, the connected units below the word midline are deleted. This step 620 is designed to delete "subscripts", such as commas and periods.

Then, MTop (mean of top) is calculated from the remaining connected units, and in step 630 of FIG. 8 , the connected units above the word midline are deleted. This step 630 is designed to remove "superscript points", such as quotation marks and the like.

MBot (average of bottom) is then computed from the remaining connected units, and then the average bottom value is computed from the remaining connected units whose bottoms are not larger than Mbot to obtain a baseline.

Furthermore, an average top value is calculated based on the obtained MTop from the remaining connected units whose top is not smaller than the MTop to obtain the upper line.

After calculating the word upper line in step 500 of accompanying drawing 1 or in step 600 of accompanying drawing 1, in step 700 of accompanying drawing 1 according to the draft word upper line for the final inspection of line information and final word interval type mark The detailed word interval type is marked.

In step 800 of accompanying drawing 1, it is judged whether the final line information is correct according to the detailed section type:

·When the interval type is 3 (it cannot be 2 to 5 at the same time), the word online must meet the following criteria:

((z1-1)/(z2+1)<0.15) and (z3/(z2+1)<0.83).

Otherwise, the line information of the word is wrong.

·When the interval type is 2 or 5, the word online must meet the following criteria:

((z1-1)/(z2+1)<0.83) and (z3/(z2+1)<0.83).

Otherwise the line information of the word is wrong, and the interval type and X height are unknown.

here,

z1: the height of the upper section of the word

z2: the height of the interval in the middle of the word

z3: the height of the lower part of the word

Zone type: Refers to the types listed in Table II.

In step 900 of the accompanying drawing 1, the final word interval type is marked according to Table VI.

Table VI

No. rule The final range type 1 When the word X height value can be calculated, the interval type is No. 2 or No. 5, and it must be No. 3 at the same time full word 2 When the word X height value can be calculated, the interval type is No. 2 or No. 5, and not 3 at the same time ascending letter words 3 When the word X height value can be calculated, the interval type is No. 3, and it is not 2 words or 3 descender word 4 The word topline is near the top of the line or the aspect ratio of the word is less than 0.65 capitalized words or numbers 5 other situations Can be an X height word

here,

Interval type: Interval type No.1 refers to the interval type of X-height words; Interval type No.2 refers to the interval type of ascending letters or full words; Interval type No.3 refers to the interval type of descending letters or full words; Interval type No. 4 refers to the interval type of words (X-height-plus word) whose upper line is close to the upper line of the line; interval type No.5 refers to the interval type of words higher than X height; and interval type No.6 refers to unknown word.

Application of Embodiment 1

Both the calculated X-height value and the identified interval type of an English word (ascender, descender, full height or X-height) are very useful information in the OCRh and OFR processes.

For example, the identified interval types can be used for dictionary classification (according to interval types) for English OCR. It can improve the speed of the ICR engine by reducing the pattern matching time, because there are fewer candidate characters in the dictionary. It also benefits accuracy, since a character is never mistakenly identified as a candidate for another interval type. Accompanying drawing 9 is shown as a kind of application of the identified interval type obtained according to the present embodiment, and it judges whether the English character image is the X height interval type, the ascending letter interval type, the descending letter interval type and the full height interval type, otherwise Dictionary identification with all range types, from which the resulting character codes and trusted values are obtained.

Interval type information is also available in OCR post-processing, which can be used to correct confusion of uppercase and corresponding lowercase characters.

In English OCR character segmentation, interval type information can be used to determine whether the "separation path" (one of the possible character separation results) is wrong.

Regarding the calculated X-height values of English words, the X-height values are very useful in the normalization of English words in both OCR and OFR processes. English words of different sizes can be normalized to a specified height according to the X-height value. Therefore features of the same level (both OCR and OFR) can be extracted from the normalized image to avoid the effect of font size.

The X height value is also important in the font size recognition of English words, and the word or line height cannot be used directly (in font size recognition), because it is impossible for characters in a word or line to occupy all 3 intervals. Accompanying drawing 10 shows an application of the X height value of an English word obtained according to this embodiment. The word X-height, image resolution, and recognized type can be used to obtain a list of X-heights of different font sizes under the specified image resolution and glyph, and then match the word X-height with the list of X-heights of different font sizes, thus obtaining font size.

This example also evaluates the accuracy of the interval type identification method on lines of printed text. Out of a total of 3302 characters, the accuracy of interval type identification is 99.67%, and the evaluation samples are printed text lines with 3 different fonts.

In short, the present embodiment provides a method for identifying the interval type of an input text line and a method for calculating the X height of a word in an input text line, both of which include: calculating line information of an input text line using a projection method Line information calculation step; the word information calculation step of using projection method to calculate the word information of the selected word in the text line; the reliability judgment step of judging whether the word information is reliable; marking the interval type of the word according to the line information and word information The interval type marking step; wherein if it is judged that the calculated word information is unreliable in the reliable judgment step, the word baseline and upper line of the selected word are calculated using the connected unit method.

This embodiment combines the projection method and the connected unit (CC) method in the row information extraction process, wherein the projection method and the connected unit method are specially designed to obtain better results, and fully consider and limit the skew, noise and so on. In this embodiment, very strict checks and verifications are performed several times to make the calculated words more and more accurate on the line. Specifically, use line baseline and projection method together in word baseline calculation to obtain accurate word baseline (refer to step 300); flexibly use the projection histogram in the designated area to calculate the upper line of short word (refer to in step 450, Step 500 and the judgment criteria in accompanying drawings 5, 6 and 7); in the new connected unit (CC) method, connected units with inappropriate areas or positions are regarded as noise or punctuation, and are eliminated to obtain more Accurate word baseline and upper line; and using the connected unit method when the reliable line information of the word cannot be obtained by the projection method (refer to step 600 and accompanying drawing 8). In addition, the detailed word interval type is marked according to the final check for line information and the draft word on-line marking the final word interval type (refer to steps 380, 700, 900 and Figures 3 and 4).

Therefore, this embodiment can accurately calculate the X-height value of an English word, and reliably identify the interval type (ascending letter, descending letter, full height or X-height) of an English word.

[Example 2]

The second embodiment pertains to a priori font recognition at the word level. It uses interval type information to classify English words into four types, each type has a different dictionary. It can identify font information (glyph, serif, point, slope and spacing) of English words with higher accuracy and speed.

It supports recognition of at least 10 glyphs, much more than current popular OCR software with document layout recovery function, such as Omnipage, FineReader, etc.

Accompanying drawing 11 shows the main flowchart of embodiment 2.

In step 100 of accompanying drawing 11, the height of each word image is normalized to WordHeight (here WordHeight=35) by bilinear interpolation, so that any size word can be processed. This is better than normalizing according to the X height; because sometimes the X height value cannot be obtained very accurately, so it will affect the subsequent feature extraction and separation processing.

Step 200 of FIG. 11 extracts font features. In principle, any global texture feature can be used in this step. Here composite 2D isotropic Gabor filters or other global features are used to extract font features from normalized word images. The 2D isotropic Gabor filter is well known in the art, and CN1271140A has provided the use of this Gabor filter to extract the features of texture images, and the details will be described below.

Specifically, 12 isotropic Gabor filters with 6 angles (0, 30, 60, 90, 120, 150 degrees) and two frequency values (0.14, 0.11) were used. In consideration of speed and accuracy, this embodiment divides each composite Gabor filter image into 2 images according to "real part" and "imaginary part". The mean and deviation are then calculated from each section separately. Thus 48 features are extracted and form a 48-dimensional feature vector (12x2x2).

Step 300 of accompanying drawing 11 judges the section type of word.

According to the interval occupied by each letter in a word, words can be classified into four interval types (X height words, ascending letter words, descending letter words and full height words).

Usually English words in different intervals have different texture features because of different combinations of characters (strokes) in different intervals, so it is proposed to use different dictionaries containing words of different interval types. The inventors have found through experiments that descenders and ascenders can be reduced to the same type because the height of the upper and lower intervals is approximately the same. The inventors also found that all capitalized words have different texture features from other ascender words, so this embodiment separates them from ascender words.

Finally, there are four dictionaries, a dictionary for X-height words, a dictionary for ascenders and descenders, a dictionary for full-height words, and a dictionary for all-caps words.

All capitalized words do not usually appear in English text, and it is very difficult to tell whether a word is all capitalized without knowing the details of each letter in the word. Therefore, under normal circumstances, this process can sometimes only obtain the interval types of ascenders/descenders, full-height or X-height words. It is not always possible to obtain interval types for English words, especially when the words are very short.

FIG. 12 is a flowchart showing an extension of step 400 of FIG. 11 .

If the interval type of the word is known, only one recognition with the known interval type is required. There are at least 40 detailed candidate glyphs (10 glyphs x 2 points x 2 slopes) in each dictionary of the specified interval type. The feature vectors extracted from the normalized word images are in turn matched against each feature vector of the detailed candidate glyphs in the dictionary of the specified word interval type. In addition, calculate a distance value (distance value), which represents the mathematical distance between the feature vector extracted from the normalized word image and the feature vector of the detailed candidate glyph in a dictionary of the specified word interval type .

A Bayesian classifier is employed in this classification step, and other classifiers such as minimum distance classification are applicable. These classifications, including Bayesian classification, are well known in the art and will not be described in detail here.

In step 500 of FIG. 11, the result of step 400 is verified. D0 is the distance value from the Bayesian classifier. TH0 (=-480) is an empirical value obtained through experiments. If the resulting distance value D0 is less than TH0, the result is reliable. Otherwise, it is determined that the word interval type may be incorrect, and step 600 is performed to identify and prioritize with multiple dictionaries.

FIG. 13 is an extended flowchart of step 600 in FIG. 11 .

This step, as shown in FIG. 13 , handles situations such as the word interval type being unavailable or possibly incorrect, including the determination that the interval type is unknown in step 300 and the result of step 400 not being available in step 500. been verified. If the tops of all letters in a word are at the same height, it will be difficult to distinguish the top line from the upper line. Similarly, if the bases of all letters in a word also have the same height, it becomes difficult to distinguish the baseline from the bottom line. Therefore, generally, compared with the interval types of ascender and descender words, it is more difficult to obtain the interval type of X-height/all ascender words, and it is easiest to obtain the interval type of full-height words. Therefore, in this step, the order of priority for dictionary selection is as shown in FIG. 13 .

The distance values D1, D2, D3, D4 are all distance values from the Bayesian classifier and are calculated to represent the feature vector extracted from the normalized word image and the detailed candidate glyphs in the standard dictionary of the four word interval types The mathematical distance between feature vectors, i.e. the distance value. TH1 (=-500) is an empirical value obtained from experiments. If Di(i=1, 2, 3, 4) is less than TH1, the result of recognition by the corresponding dictionary is very reliable, then we will not continue to use other dictionaries for recognition, and the result will be output as the final result. Otherwise, the distance values D1, D2, D3, D4 are compared with each other to obtain the minimum distance. The corresponding glyph is the final detailed glyph (selected from 40).

Application of Embodiment 2

The methods described above can be used for page layout analysis and recovery. It identifies font information for each word in an image of an English document.

It can also be used in full font OCR system. Font information of English words can be predicted and used to select an OCR dictionary with a specified font to improve OCR accuracy.

The above flow chart is simply shown in Figure 14: block selection, line segmentation, and word segmentation are performed on the English document image in order to obtain word images and implement word interval type identification, and then perform the processing defined by this embodiment. The obtained font information for each word can be used for layout restoration, single-font OCR, etc.

This embodiment compares Embodiment 2 with popular OCR software with font recognition (layout recovery) on the market. The accuracy of Example 2 is much higher than them.

Various accuracies of glyph recognition are tabulated in Table VII.

Omnipage11.0 FineReader5.0 TypeReaderProfessional6.0 TextBridgeProMillennium Example 2 Benchmark 1 46.24% 30.09% 30.39% 27.75% 95.57% Benchmark 2 33.86% 20.36% 28.96% 22.98% 88.73% Benchmark 3 30.49% 10.09% 19.49% 14.52% 95.24%

Benchmark 1: Lines of printed text with uniform fonts for all words (820 lines, 93344 words in total).

Benchmark 2: Lines of printed text (1560 lines, 18009 words in total) with 2 different fonts.

Benchmark 3: Lines of printed text (288 lines, 3637 words in total) with 3 different fonts.

In short, this embodiment provides a font recognition method, including: a normalization step of normalizing the words of the input image to a predetermined height; a feature extraction step of extracting features from the normalized words; judging the A judgment step of section type; a classification step, and a recognition step of recognizing font information of a word based on a result of the classification step.

This font recognition method normalizes the height of the word image to a given height, thereby extracting the features of the word image with the same interval type at the same level, because in font recognition, the dictionary is evaluated according to the interval type of the word Classification (refer to step 100 in Figure 11). First consider the word interval type to select the corresponding "interval type" dictionary in the grapheme classification (refer to accompanying drawing 12). When the resulting distance is greater than a given threshold or the word interval type is unknown, a priori arranged multi-dictionary recognition process will be used (refer to FIG. 13 ). Therefore, such an embodiment will take full advantage of the information about the interval type, and the overall OCR process becomes more efficient.

[Example 3]

Embodiment 3 adopts the word pair mechanism in the OFR process of the line image. The word pair mechanism is based on the font classification method of English words and simultaneously considers the characteristics of the font distribution in the actual English text. It uses a two-stage result conditioning technique based on contextual font information, which enables greater precision and neater output within lines of text. An accurate size recognition method is used in this embodiment.

It supports the recognition of 10 glyphs, much more than ordinary OCR software (such as Omnipage, FineReader, etc.) with file layout recovery function. It can achieve much higher accuracy (both font style and font size) than other software.

Accompanying drawing 15 shows the main flowchart of this embodiment 3.

In step 100 of FIG. 15, the X-height values and interval types of all words in the row are first calculated. Preferably use the X height value and the interval type of the words obtained by the method described in the above-mentioned embodiment 1, or use those methods in the prior art, such as the character height distribution histogram method disclosed in US5883974 and for example in Optical fontrecognition from The vertical projection distribution method disclosed by projection profiles (Electronic publishing, VOL.6(3), 249-260 (September 1993)).

In step 200 of FIG. 15 , this embodiment proposes to divide all the words in the row into word pairs starting from the first word in the row.

Through experiments, it is known that the recognition results of longer words (with larger word width) are more reliable than those of shorter words.

Each word pair consists of two adjacent words (if the number of words in the row is odd, the last word pair consists of only one word). The fonts for longer and shorter words will be identified in different ways below.

Printed inIndia)中公开的基于词的“英文单词识别”的方法或者如在US6,337924和US6,496,600中公开的后验字体识别法。In step 300 of FIG. 15, the longer word of each word pair can be identified by various methods using font classification of English words. These methods include the method disclosed in Example 2 above, as described in Script identification in printed bilingual documents to DDHANYA, A G RAMAKRISHNAN_ and PEETA BASA PATI (Sadhana Vol.27, Part 1, February 2002, pp.73-82.

The word-based "English word recognition" method disclosed in Printed in India) or the a posteriori font recognition method as disclosed in US6,337924 and US6,496,600.

In step 400 of FIG. 15, the font of the shorter word in each word pair is identified.

In this step, there are two facts that must be considered first.

1) The actual text line, if the detailed glyphs of the two longer words around (left and right) the shorter word are the same, then the detailed glyphs of the shorter word are also very likely to be the same.

2) As mentioned in step 200, the font recognition results of longer words (with longer word width) are more reliable than the font recognition of shorter words in the English word font classification.

Considering these facts, the flowchart shown in accompanying drawing 16 is used to carry out the step 400 of accompanying drawing 15.

In step 410 of the accompanying drawing 16, first judge whether the word to be recognized is at the edge of the text line? If not, there must be two longer words surrounding the word to be recognized.

From step 420 to step 460 of FIG. 16, the detailed glyphs of two longer words around the word to be recognized are obtained and compared with each other. (That is, glyph 1 is set by the glyph of the longer word in the same word pair, and glyph 2 is set by the glyph of the longer word in the previous or next word pair.)

At step 420 it is determined whether the shorter word in the current pair is the first word in the pair. If the decision result of step 420 is yes, then the detailed font of the longer word in the last word pair is set as " font 2 ", otherwise in step 440 the detailed font of the longer word in the next word pair is set as "Glyph 2".

In step 450, the detailed grapheme of the longer word in the current word pair is set as "grapheme 1".

In step 460 it is determined whether the set "Font 1" is equal to the set "Font 2". If the detailed grapheme of the longer word is the same, then we don't have to identify the shorter word in the current word pair, whose detailed grapheme is also the same "grapheme 1" (step 480). Otherwise, the shorter word of the current word pair needs to be identified by using the font classification of the English word (step 470). This method has been introduced when describing step 300, so their description is omitted here.

Through this flowchart, font identification of the shorter word in each word pair can achieve higher accuracy, higher speed, and cleaner output.

Now that the font results for each word have been obtained, step 500 of FIG. 15 begins to adjust the font results based on the font information of adjacent words.

Figure 17 is an extended flow chart of step 500 in Figure 15 .

Three candidate fonts are extracted for each word in order of similarity by comparing the font result of each word with a dictionary containing multiple fonts of the standard pattern, where the first candidate font has the largest relative to the font result of each word Similarities are matched, so corresponding distance values are obtained. Here, three candidate fonts are preferred by practice. However, according to actual requirements, two candidate fonts or more than three candidate fonts may also be used in this embodiment.

In step 510 of FIG. 17, the distribution of detailed glyphs in the row is counted according to the first candidate font for the word.

Then loop through the remaining steps for each word in the line.

When the current word is the first or last word in the row, the processing of this embodiment will judge whether the detailed fonts of the row are roughly consistent (step 520 of accompanying drawing 17). Its standard is that more than 80% of the words in the row have the same detailed font at any time or when the number of words in the row is no more than 6, more than 60% of the words in the row have the same detailed font. If the detailed glyphs of the row are substantially uneven, the processing of this embodiment does not perform adjustments to the current word and continues the loop processing for other words in the row.

If in step 520 the detailed font of the row is roughly consistent, then go to step 521 of accompanying drawing 17 to judge whether the first candidate font of current word is consistent with the main font of this row. If step 521 judges that the first candidate font of the current word is consistent with the main glyph of the row, a loop process is executed for other words in the row.

If in step 521 it is judged that the first candidate font of the current word is inconsistent with the main font of the row, then go to step 522 of accompanying drawing 17 to judge whether the second candidate font of the current word is consistent with the main font of the row and corresponding Whether the distance also satisfies the specified condition. If it is judged that the second candidate font of the current word is consistent with the main font of the row and the corresponding distance also satisfies the specified condition, then the current second candidate font is exchanged with the first candidate font, otherwise go to step 523 of accompanying drawing 17 To determine whether the current third candidate font is consistent with the main glyph of the line and whether the corresponding distance also meets the specified condition.

If step 523 judges that the current third candidate font is consistent with the main glyph of the row and the corresponding distance also satisfies the specified condition, then exchange the current third candidate font and the first candidate font, otherwise go to step 530 of accompanying drawing 17 To judge whether its candidate fonts are reliable. If the candidate fonts are unreliable when it is determined that the distance between all the candidate fonts is large, then it is considered that the recognition process has not been completed well, and then the first candidate font is set to be the same as the candidate fonts of its adjacent words, that is, the adjacent word Font sets the first candidate font (step 540 of FIG. 17).

When the current word is not the first or last word in the row, the processing of this embodiment executes step 550 of accompanying drawing 17 to judge whether the detailed font in the row is roughly consistent and whether the first candidate font of the current word is consistent with The main glyphs of the line are consistent. The judgment condition is that more than 80% of the words in the line have the same detailed font.

Step 551 of accompanying drawing 17 further judges whether the detailed font of its adjacent 2 words is the same and whether the first candidate font is different from them.

Steps 552 and 553 of accompanying drawing 17 further judge whether the second and the third candidate fonts are consistent with 2 adjacent words and whether the distance of the candidate fonts satisfies the specified condition.

If in step 552 it is judged that the second candidate font of the current word is consistent with 2 adjacent words and the corresponding distance also satisfies the specified condition, then exchange the current second candidate font and the first candidate font, otherwise go to step 553 to Determine the current third candidate font. If judge in step 553 that the 3rd candidate font of current word is consistent with 2 adjacent words and corresponding distance also satisfies specified condition, then exchange current 3rd candidate font and first candidate font, otherwise turn to accompanying drawing 17 Step 560 to determine whether the candidate font is reliable. If the candidate fonts are unreliable when it is determined that the distances of all candidate fonts are large, it is considered that the recognition process has not been completed well, so the first candidate font is set to be the same as the candidate fonts of its adjacent words, that is, the Font sets the first candidate font (step 570 of FIG. 17).

Step 600 of FIG. 15 further adjusts the font results based on the row's font distribution information.

FIG. 18 is an extended flowchart of step 600 in FIG. 15 .

In step 610 of FIG. 18, the distribution, serif and spacing of the detailed glyphs in the line are counted according to the first candidate font of the adjusted word.

Step 620 of FIG. 18 judges whether the serifs and spacing are absolutely uniform and whether the detailed glyphs are substantially consistent in the row.

The conditions used are as follows:

All words in the line have the same serif of the first candidate font;

All words in the line have the same spacing in the first candidate font;

More than 75% but not 100% of the words in the line have the same detailed glyph.

When all three conditions are met, set the first candidate font in the line with the main detailed glyph and use the first candidate font for each word as the recognized font result. But when one of the above three conditions is not satisfied, the result of step 620 is no, then the processing of this embodiment does not perform adjustment and instead uses the current first candidate font of each word as the recognized font result.

In step 700 of FIG. 15, the font size is calculated from the image resolution, recognized glyph and word X height (or word interval type).

FIG. 19 is an expanded flowchart of step 700 in FIG. 15 .

If the X height of the word is available (step 100 of Figure 15), then the left branch will be executed.

In step 720 of accompanying drawing 19, " image resolution/font style/font size/X height " table in the font size dictionary of querying by known input image resolution and recognized font shape when inputting, and obtain the X of different font sizes height list.

In step 730 of FIG. 19, the nearest X-height in the list is searched based on the X-height of the input word. The X-height with the smallest |xx _i |/ _xi in the list is the closest one. The corresponding font size is the recognized size. Here "x" refers to the value of the input X height, and "xi" refers to the value of the ith X height in the list.

Otherwise, if the word's X-height is not available or accurate, the right branch is taken.

Data 740 is the word interval type (of the selected dictionary) from the grapheme class.

Steps 750 and 760 of FIG. 19 are similar to steps 720 and 730 of FIG. 19. The only difference is that they are based on word height, not X height. Step 750 of accompanying drawing 19 includes the prior font size dictionary of " image resolution/font style/font size/word height " table through known input image resolution and recognized font form query when inputting, and obtains the word height of different font sizes list; step 760 of accompanying drawing 19 matches search for the nearest word height in the word height list of the input word; and

Identify the corresponding font size as the identified size.

Application of Embodiment 3

The method described in Example 3 can be used for page layout analysis and recovery. It can identify font information of English text lines with higher accuracy and speed.

It can also be used in full font OCR system. Font information can be predicted and used to select an OCR dictionary with a specified font to improve OCR accuracy.

The flow chart can be simplified as shown in FIG. 20: perform block selection on an English document image, line segmentation to obtain a line image, and implement processing as defined in this embodiment. The obtained font information can be used for layout recovery, single-font OCR, etc.

The applicant compared this embodiment with popular OCR software with font recognition (layout recovery) functions on the market. The accuracy of this embodiment is much higher than them in font style and font size. In addition, the applicant also compared this embodiment with the above-mentioned embodiment 2 in which only a single English word font classification is used to replace steps 200-600. This embodiment improves both accuracy and speed.

Table VIII shows the comparison results of the accuracy of glyph recognition:

Omnipage11.0 FineReader5.0 TypeReaderProfessional6.0 TextBridgeProMillennium Example 2 Example 3 Benchmark 1 46.24% 30.09% 30.39% 27.75% 95.57% 99.16% Benchmark 2 33.86% 20.36% 28.96% 22.98% 88.73% 90.33% Benchmark 3 30.49% 10.09% 19.49% 14.52% 95.24% 97.09%

Table IX shows the comparison results of the accuracy of font size recognition:

Omnipage11.0 FineReader5.0 TypeReaderProfessional6.0 TextBridgeProMillennium Example 2 Example 3 Benchmark 1 45.83% 4.65% 38.96% 15.22% 61.25% 61.81% Benchmark 2 40.42% 10.54% 57.79% 18.56% 81.97% 82.93% Benchmark 3 21.14% 20.76% 23.62% 10.61% 76.02% 76.82%

The speed of this embodiment is 28.9% higher than that of the above-mentioned embodiment 2 in terms of "baseline detection program speed".

Benchmark 1: Lines of printed text with uniform font for all words (8020 lines, 93344 words in total)

Benchmark 2: Lines of printed text with 2 different fonts (1560 lines, 18009 words in total)

Benchmark 3: Lines of printed text with 3 different fonts (288 lines, 3637 words in total)

Benchmarking Program Speed: Selected lines of printed text from Benchmarking Programs 1-3 (201 lines, 2659 words total)

In short, this embodiment introduces a method for optical font recognition, including: a division step of dividing words of an input text image into word pairs by lines; a font recognition step of identifying font information of longer words in each word pair; based on The font information of the longer word in the word pair including the shorter word and the font information of the longer word in the word pair adjacent to the shorter word identify the font of the font information of the shorter word in each word pair A discriminating step; a fine adjustment step of adjusting font information of a word according to font information of adjacent words; a coarse adjustment step of adjusting font information of a row according to distribution of font information in the row; and a recognition step of identifying a font size of a word of a row.

The present embodiment introduces the word pair mechanism in the line font recognition process, and this word pair mechanism has considered the font recognition feature of the word level and the characteristics of the font distribution in the actual English text (referring to steps 200, 300 and accompanying drawing 16) . This embodiment also introduces post-processing that employs a two-level (based on font information and font distribution information, respectively, of the adjacent words of the line) result conditioning technique that enables higher precision within the text line And more tidy output (refer to accompanying drawing 17 and accompanying drawing 18). This embodiment also adopts the precise size identification technology to combine "interval type information" and "X height value" (refer to accompanying drawing 19). Therefore, this embodiment can achieve higher accuracy and neater output in the text line, and the method introduced in this embodiment supports the recognition of 10 glyphs, which is far more than the popular one with the file layout recovery function OCR software, such as Omnipage, FineReader, etc., and can achieve much higher accuracy (both font style and font size) than other software.

[Example 4]

As described above, the first, second and third embodiments may be used alone, or may be used in combination. Fig. 21 is a flow chart for optical character recognition according to the fourth embodiment of the present invention, wherein the above three embodiments can be applied.

Embodiment 1 can be used for section type identification and X-height calculation of a word image. In step 100, the word X-height and section type are calculated using Example 1. They are very useful information in glyph classification and font recognition.

Steps 200, 300, 400 relate to word image normalization, font feature extraction and glyph classification already described in Embodiment 2. Detailed glyphs (combining point and slant) are available after glyph classification, and they can also be used (together with updated word interval types) for font size recognition, eg step 600 . At the same time, font information (font style, font size, point size, slope, spacing and serif) of English words can be obtained accurately, such as step 500 .

In addition, embodiments 1-4 are not limited to English words, based on their processing principles, these embodiments can directly relate to other Roman characters.

In the foregoing description, the present invention has been described in preferred embodiments as methods or software programs. With regard to the present invention, it is readily understood that the present invention is preferably used on any known computer system such as a personal computer. Therefore, the computer system will not be discussed in detail. It is also to be noted that the image may be input directly into the computer system (eg by a digital camera) or digitized (eg by scanning) prior to input into the computer system.

Furthermore, as used herein, a computer-readable storage medium having stored therein a computer program for performing the methods described above may include, for example, magnetic storage media such as magnetic disks (such as floppy disks) or magnetic tapes; optical storage media such as optical disks, optical tapes, or A machine-readable bar code; a solid-state electronic storage device such as random access memory (RAM); or any other physical device or medium used to store computer programs.

Furthermore, those skilled in the art readily recognize that equivalents of the software described above can also be designed in hardware.

With reference to embodiment 3, accompanying drawing 22 shows the device that is configured to implement the above-mentioned process of described optical character recognition, in brief, this device comprises:

A division device 100 for dividing the words of the input text image into word pairs by lines;

A font recognition device 101 that recognizes font information of longer words in each word pair;

Discriminating the font information of the shorter word in each word pair based on the font information of the longer word in the word pair that includes the shorter word and the font information of the longer word in the word pair adjacent to the shorter word. font identification device 102;

A fine adjustment device 103 for adjusting the font information of words according to the font information of adjacent words; and

coarse adjustment means 104 for adjusting the font information of the row according to the distribution of the font information in the row; and

Recognition means 105 for recognizing the font size of the words of the line.

Referring to Embodiment 2, accompanying drawing 23 shows the device configured to implement the above-mentioned process for font recognition described, in short, the device includes:

A normalization device 200 for normalizing the words of the input image to a predetermined height;

A feature extraction device 201 for extracting features from normalized words;

A judging device 202 for judging the interval type of a word;

sorting means 203, and

Recognition means 204 that recognize font information of words based on the results obtained by the classification means.

The classification device 203 further includes a comparison device that compares the extracted features of the normalized word with the features of the candidate glyphs in the dictionary of the judged word interval type, and an acquisition device that obtains a distance value from the comparison device, so that the font The apparatus of the method of identification may comprise, in a step, verification means for verifying the obtained distance value with a predetermined threshold value. If the checking means determines that the distance value is less than a predetermined threshold, then the fine-graining means uses a plurality of dictionaries, the fine-graining means being configured to;

comparing the extracted features of the normalized word with the features of the candidate glyphs in the dictionary of at least one other word interval type except the judged word interval type,

obtain distance values from at least one comparison step, respectively, and

The distance value is verified against another predetermined threshold to identify the font information of the word based on the result of the verification step.

In addition, if the judging means 202 cannot successfully judge the interval type of the word, the subdivided classifying means is used.

Considering the foregoing, accompanying drawing 23 shows a modified device for implementing the font recognition method, including: a normalization device 300 that normalizes the input image to a predetermined height; a feature extraction device that extracts features from the normalized word 301; subdivided classification means 302 directly using a plurality of dictionaries; and recognition means 303 for identifying font information of words based on results obtained by the subdivided classification means.

With reference to Embodiment 1, accompanying drawing 25 shows the device that is configured to implement all the above-mentioned processes of the described interval type of discriminating input text line, in brief, this device comprises:

A line information calculation device 400 for calculating line information of an input text line using a projection method;

A first word information calculating means 401 for calculating the word information of the selected word in the text line using a projection method;

A reliability judging device 402 for judging whether word information is reliable;

By using the connected unit method to calculate the word base line and the upper line of the word baseline and the upper line of the selected word that is not judged as reliable by the reliability judging means; and

The interval type marking means 404 for marking the interval type of the word according to the line information and the word information.

The device for discriminating the interval type of the input text line may further include judging means for judging whether the selected words in the text line are short words, and the first word information calculation means 401 calculates separately by using the projection method in the designated area of the word interval The word information of the judged short word.

Furthermore, the means for discriminating the section type of an input text line can obviously be modified as a means for calculating the X-height of a word in an input text line by using the section type obtained by said means for discriminating the section type of an input text line.

The invention has been described with reference to specific embodiments. It should be understood that the present invention is not limited to the above description, and those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention.

Claims (25)

1. A method for optical character recognition, comprising the steps of: The step of dividing the words of the input text line image into a plurality of word pairs, wherein each of the plurality of word pairs includes two adjacent words; A font recognition step of identifying the font information of the longer long word in two adjacent words included in a word pair by using the English word font classification method; A font identification step of discriminating the font information of the short word based on the font information of the long word in the word pair including the short word and the font information of the long word in another word pair adjacent to the short word, wherein, A short word is the shorter of two adjacent words included in a word pair, Wherein said font identification step comprises: comparing the font information of the long word in the other word pair adjacent to the short word to be identified with the font information of the long word in the word pair that includes the short word to be identified; and If the font information of the long word in the other word pair adjacent to the short word to be identified is identical to the font information of the long word in the word pair that includes the short word to be identified, then determine the word to be identified in the word pair The identified short word has the same font information as the long word in the word pair and the long word in the adjacent word pair; If the font information of the long word in the other word pair adjacent to the short word to be identified is different from the font information of the long word in the word pair that includes the short word to be identified, then by using the English word font classification method Font information identifying the phrase; and If the short word to be identified in the word pair is the first or last word in the image of the input text line, the font information of the short word is identified by using the English word font classification method. 2. The optical font recognition method according to claim 1, further comprising a fine adjustment step of adjusting the font information of the word according to the font information of adjacent words, wherein the fine adjustment step is performed after the font identification step. 3. The optical font recognition method according to claim 2, wherein the fine adjustment step of adjusting font information comprises: determining whether a first candidate font for each word in the input text line image is available; If it is determined that the first candidate font of the word is not available, then determine whether the second candidate font of the word is available; If it is determined that the second candidate font of the word is available, the second candidate font of the word is exchanged with the first candidate font of the word. 4. The optical font recognition method according to claim 3, wherein the fine-tuning step of adjusting the font information further comprises: if determining that the second candidate font of the word is unavailable then determining whether the third candidate font of the word is available, and If it is determined that the third candidate font of the word is available, the third candidate font of the word is exchanged with the first candidate font of the word. 5. The optical font recognition method according to claim 4, wherein if it is determined that the third candidate font of the word is unavailable, then it is judged whether all three candidate fonts of the word are reliable; and If it is determined that all three candidate fonts for the word are unreliable then the first candidate font for the word is set in the font of the adjacent word. 6. The optical font recognition method according to claim 5, wherein the fine adjustment step comprises the following preprocessing process: comparing the font information of each word in the input text line image with a dictionary containing the font information of a plurality of fonts; and The order of similarity obtains three candidate fonts and three corresponding distance values for the word, The condition for judging that all three candidate fonts of the word are unreliable includes that all three distance values are greater than a predetermined threshold. 7. The optical font recognition method according to claim 3, wherein the fine adjustment step comprises the following preprocessing process: comparing the font information of each word in the input text line image with a dictionary containing the font information of a plurality of fonts; Obtaining at least first and second candidate fonts and at least two corresponding distance values for the word in order of similarity; and The distribution of detailed glyphs in the input text line image is counted according to the first candidate font for the word. 8. The optical font recognition method according to claim 7, wherein for the first or last word in the input text line image, determining the available condition of the first candidate font of the word comprises: Detailed glyphs are not consistent across input text line images, or The first candidate font of the word is consistent with the main glyph of the input text line image. 9. The optical font recognition method according to claim 7, wherein for the first or last word in the input text line image, determining the condition that the second candidate font of the word is available comprises: The second candidate font of the word is consistent with the main glyph of the input text line image and the corresponding distance value is smaller than a predetermined threshold. 10. The optical font recognition method according to claim 7, wherein said candidate fonts comprise a first candidate font, a second candidate font and a third candidate font, For the first or last word in the input text line image, the conditions for determining that the third candidate font of the word is available include: The third candidate font of the word is consistent with the main glyph of the input text line image and the corresponding distance value is smaller than a predetermined threshold. 11. The optical font recognition method according to claim 7, wherein for each word except the first and last words in the input text line image, the condition for determining that the first candidate font of the word is available comprises : the detailed glyph is consistent in the input text line image, and the first candidate font for the current word is consistent with the main glyph in the input text line image, or The first candidate font of the current word is different from the first candidate fonts of two adjacent words in the input text line image, or the detailed fonts of the two adjacent words are the same. 12. The optical font recognition method according to claim 7, wherein for each word except the first and last words in the input text line image, the condition for determining the availability of the second candidate font for the word comprises : The second candidate font of the word is consistent with two adjacent words and the corresponding distance value is smaller than a predetermined threshold. 13. The optical font recognition method according to claim 7, wherein said candidate fonts comprise a first candidate font, a second candidate font and a third candidate font, For each word except the first and last words in the input text line image, the conditions for determining the availability of the third candidate font for the word include: The third candidate font of the word is consistent with the main glyphs of two adjacent words and the corresponding distance value is smaller than a predetermined threshold. 14. The optical font recognition method according to claim 1 or 2, further comprising a rough adjustment step of the font information of the word according to the font information distribution in the input text line image, wherein, after the font identification step, the step of performing the The rough adjustment steps described above, The coarse adjustment step of adjusting the font information of the input text line image comprises: counting at least one of serifs and spacing and distribution of detailed glyphs in the input text line image; determining whether the at least one of serifs and spacing and detailed glyphs are consistent in the input text line image, and If it is determined that the above distribution satisfies the condition, then set the first candidate fonts of all words in the input text line image with the main detailed font and use the first candidate fonts of each word as the recognized font result. 15. The optical font recognition method according to claim 1, further comprising a recognition step of recognizing a font size of a word in the input text line image, wherein said recognition step is performed after said font identification step. 16. The optical font recognition method according to claim 15, wherein the recognition step of recognizing the character size of the word in the input text row image comprises: Determine whether the input word X height of the word is available; If the input word X height is available, query the prior font size dictionary containing the "image resolution/font style/font size/X height" table with the recognized glyph of the known word and the input image resolution, and obtain a list of X heights of different font sizes ; match the word's X-height to the X-heights in the X-heights list; and Use the corresponding font size as the recognized font size. 17. The optical font recognition method according to claim 15, wherein the recognition step of recognizing the character size of the word in the input text row image comprises: If the X height of the input word is not available, query the prior font size dictionary containing the table "image resolution/font style/font size/word height" with the recognized glyph of the known word and the input image resolution, and obtain the word heights of different font sizes list; match the word height of the word with the word heights in the word height list; and Use the corresponding font size as the recognized font size. 18. A font recognition method, comprising: A normalization step that normalizes the words of the input image to a predetermined height; A feature extraction step that extracts features from normalized words; The step of judging the interval type of the judgment word; Classification step, if the interval type of the word can be obtained in the judgment step, then compare the feature of the extracted feature of the normalized word with the feature of the candidate glyph in the dictionary of the judged word interval type, and obtain the distance value according to the comparison ; a verification step of verifying the obtained distance value with a predetermined threshold; Subdividing the classification step, if the interval type of the word cannot be obtained in the judgment step or if the distance value is not less than a predetermined threshold in the verification step, then the extracted features of the normalized word are compared with the features of the candidate glyphs in a plurality of dictionaries; as well as an identifying step of identifying font information of words based on the result of the classifying step if said distance value is smaller than a predetermined threshold in the verifying step, and if the interval type of the word cannot be obtained in the judging step or if said in the verifying step If the distance value is not less than a predetermined threshold, the font information of the word is identified based on the result in the subdivision step. 19. The font recognition method according to claim 18, wherein the judging step includes a step of judging whether the interval type of the word has been known from the outside. 20. The font recognition method according to claim 18, wherein, in the classifying step, the dictionary is at least selected from a dictionary of X-height words, a dictionary of all uppercase words, ascending letter words, and descending letters based on the judged interval type. Choose from a dictionary of alphabetic words and a dictionary of full-height words. 21. The font recognition method according to claim 18, wherein each dictionary of interval types has at least 40 types of candidate glyphs. 22. The font recognition method according to claim 18, wherein the subdividing step comprises: comparing the extracted features of the normalized word with the features of the candidate glyphs in the dictionary of at least one other word interval type except the judged word interval type, obtaining a distance value based on said at least one comparison, respectively, and The distance value is verified against another predetermined threshold to identify the font information of the word based on the result of the verification step. 23. The font recognition method according to claim 22, wherein the dictionary of at least one word interval type to be compared comprises dictionaries in the following order: a dictionary of X height words, a dictionary of all uppercase words, a dictionary of ascender words and descender words Dictionary and dictionary of full height words. 24. An optical font recognition device comprising: A division means for dividing the word of the input text line image into a plurality of word pairs, wherein each of the plurality of word pairs includes two adjacent words; A font recognition device for recognizing the font information of the longer word among two adjacent words included in a word pair by using the English word font classification method; A font identifying means for discriminating font information of a short word based on font information of a long word in a word pair including the short word and font information of a long word in another word pair adjacent to the short word, wherein the short word The word is the shorter of the two adjacent words included in the word pair, Wherein said font identification device comprises: comparing the font information of the long word in the other word pair adjacent to the short word to be identified with the font information of the long word in the word pair that includes the short word to be identified; and If the font information of the long word in the other word pair adjacent to the short word to be identified is identical to the font information of the long word in the word pair that includes the short word to be identified, then determine the word to be identified in the word pair The identified short word has the same font information as the long word in the word pair and the long word in the adjacent word pair; If the font information of the long word in the other word pair adjacent to the short word to be identified is different from the font information of the long word in the word pair that includes the short word to be identified, then by using the English word font classification method Font information identifying the phrase; and If the short word to be identified in the word pair is the first or last word in the image of the input text line, the font information of the short word is identified by using the English word font classification method. 25. A font recognition device comprising: normalization means for normalizing the words of the input image to a predetermined height; A feature extraction device for extracting features from the normalized words; A judging device for judging the interval type of a word; The classification device, if the judging device can obtain the interval type of the word, then compare the extracted feature of the normalized word with the feature of the candidate glyph in the dictionary of the judged word interval type, and obtain the distance value according to the comparison; verification means for verifying the obtained distance value with a predetermined threshold; Subdivided classification means, if the judging means cannot obtain the interval type of the word or if the verification means verifies that the distance value is not less than a predetermined threshold, then the extracted features of the normalized word are compared with the features of the candidate glyphs in a plurality of dictionaries; and recognition means, if the verification means verifies that the distance value is less than a predetermined threshold, then identify the font information of the word based on the result of the classification means, and if the judging means cannot obtain the interval type of the word or if the verification means verifies that the distance value is not less than A predetermined threshold value is used to identify font information of words based on the results in the fine classification means.

Images