KR20190143525A - A Method of Detecting Character Data through a Adaboost Learning Method - Google Patents

Abstract

Extracting candidate regions of characters from MV features and AdaBoost methods from input images, and detecting only real problem areas except non-character regions from extracted character regions using geometric features A method of detecting character data using boost learning, comprising: (a) extracting a Modified Census Transform (MCT) feature from the input image; (b) detecting candidate regions of a text area in the input image by inputting the extracted MCT feature to the text detection classifier using a text detection classifier generated by an Adaboost learning method; And (c) using the geometric features of the candidate region to filter the candidate regions to detect the final character region, thereby using the MCT feature and the AdaBoost method as well as the geometric features as candidates. By effectively filtering the area, it is possible to detect the character areas more accurately than in the prior art.

Description

A method of Detecting Character Data through a Adaboost Learning Method}

The present invention relates to a method of detecting text data using Adaboost learning, which accurately detects only a text area using an MCT feature and an Adaboost method from an input image.

In addition, the present invention extracts candidate regions of characters from the input image by using MCT features and Adaboost method, and removes only non-character regions from candidate regions of characters extracted by using geometric features. And character data detection method using Adaboost learning.

Recently, due to the remarkable development of hardware and software, the amount of various kinds of image data collected from smart phones, CCTVs, black boxes, drones, satellites, and digital cameras is increasing exponentially. In general, image big data is recognized as a next-generation new technology that seeks to discover new values by extracting meaningful information by analyzing this type of atypical image data, which has not been properly analyzed in the past, instead of analyzing existing standardized data. [Non Patent Literature 1].

It is very important to analyze various kinds of image data and extract only the character region included in the image [Non-Patent Document 2]. This is because the characters included in the image can provide important key information representing the meaning of the image. Extraction of license plates of automobiles using image analysis is particularly important in the field of character extraction, which is a necessary task for access control of parking lots and cracking down on speeding vehicles on the road [Non-Patent Document 3].

Existing researches that automatically extract only the text area from various kinds of input images can be found in the related literature. In [Non-Patent Document 4], input images are divided in units of N × N square blocks, and character regions are extracted by using a feature that a high frequency component in a horizontal or vertical direction is relatively large in a region in which characters exist. In [Non-Patent Document 5], using a saturation data, a color transition map between a background area and a text area was created to detect a text area from an input image. In [Non-Patent Document 6], a license plate of a vehicle or a two-wheeled vehicle located in front of the vehicle was detected by detecting and tracking an object using a gray image captured by a fixed camera. In [Non-Patent Document 7], a license plate area of a vehicle was detected using yellow or green color information from an image. However, this method has the disadvantage that the background should be a relatively simple image. In addition to the methods described above, many methods related to extracting text areas continue to be introduced in the literature.

The existing character area detection methods described above can also detect the character areas to some extent. However, there is a constraint that the accuracy can be guaranteed to some extent only in a predetermined constant environment.

Therefore, there is a need for a technology capable of accurately detecting characters in an image regardless of the surrounding environment.

 J.-G. Ko, Y.-S. Bae, J.-Y. Park, and K. Park, "Technologies Trends in Image Big Data Analysis," Electronics and Telecommunications Trends, Vol. 29, No. 4, pp. 21-29, August 2014.  S.-W. Jang, K.-J. Ahn, and G.-Y. Kim, "Robust Caption Segmentation Using DCT Coefficients and Edges," Journal of Korean Institute of Next Generation Computing, Vol. 11, No. 3, pp. 63-72, June 2015.  K. Lim, H. Byun, and Y. Choi, "Vehicle License Plate Detection in Road Images," Journal of Korean Institute of Information Scientists and Engineers, Vol. 43, no. 2, pp. 186-195, February 2016.  X. Qian, G. Liu, H. Wang, and R. Su, "Text Detection, Localization, and Tracking in Compressed Video," Signal Processing: Image Communication, Vol. 22, no. 9. pp. 752-768. October 2007.  W. Kim and C. Kim, "A New Approach for Overlay Text Detection and Extraction from Complex Video Scene," IEEE Transactions on Image Processing, Vol. 18, No. 2, pp. 401-411, February 2009.  H. Lee, S. Chen, and S. Wang, "Extraction and recognition of license plates of motorcycles and vehicles," In Proc. of the 17th IEEE International Conference on Pattern Recognition, pp. 356-359, August 2004.  K. Deb and K. Jo, "HSI Color-based Vehicle License Plate Detection," In Proc. of the International Conference on Control, Automation and Systems, pp. 687-691, 2008.  B. Froba and A. Ernst, "Face Detection with the Modified Census Transform," In Proc. of the IEEE International Conference on Automatic Face and Gesture Recognition, pp. 91-96, May 2004.  H.-W. Jo and B. Moon, "A Modified Census Transform Using the Representative Intensity Values," In Proc. of the International SoC Design Conference (ISOCC), pp. 309-310, November 2015.  C. Gao, P. Li, Y. Zhang, J. Liu, and L. Wang, "People Counting Based on Head Detection Combining Adaboost and CNN in Crowded Surveillance Environment," Neurocomputing, Vol. 208, pp. 108-116, October 2016.  P. Viola and M. Jones, "Rapid Object Detection Using a Boosted Cascade of Simple Features," In Proc. of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Vol. 1, pp. 511-518, December, 2001.  E. Rosten and T. Drummond, "Machine Learning for High-Speed Corner Detection," In Proc. of the European Conference on Computer Vision, pp. 430-443, May 2006.

SUMMARY OF THE INVENTION An object of the present invention is to solve the problems described above, extracts candidate regions of a character from an input image using MCT features and adaptive boosting (Adaboost) learning method, and extracts extracted candidate character regions. The present invention provides a method of detecting character data using Adaboost learning, which robustly detects only actual character regions excluding non-character regions by filtering using geometric information.

In order to achieve the above object, the present invention relates to a method of detecting character data using Adaboost learning, which detects a character region in an input image, comprising: (a) extracting a Modified Census Transform (MCT) feature from the input image; (b) detecting candidate regions of a text area in the input image by inputting the extracted MCT feature to the text detection classifier using a text detection classifier generated by an Adaboost learning method; And (c) filtering the candidate regions to detect a final character region using geometric features of the candidate regions.

In addition, the present invention is characterized in that in the character data detection system using Adaboo boost learning, MCT feature Γ (x) of the input image is extracted by the following equation (1).

[Equation 1]

는 커널 내에 위치한 화소들의 평균 화소 값을 나타내고, n'은 커널의 중심과 인접한 화소들의 집합을 나타내고, ζ()는 비교 함수로서 I(y)가 평균 화소 값 보다 크면 1을 출력하고, 그렇지 않으면 0을 출력하는 함수이고,

는 10진수 변환 연산자로서 비교 함수 ζ()의 결과로 나오는 아홉 자리의 2진수 배열을 10진수로 변경해 주는 연산자임.Where I (x) represents the pixel value of x,

Denotes the average pixel value of the pixels located in the kernel, n 'denotes the set of pixels adjacent to the center of the kernel, ζ () is a comparison function, and outputs 1 if I (y) is greater than the average pixel value; Is a function that outputs zero,

Is a decimal conversion operator that converts the nine-digit binary array resulting from the comparison function ζ () into a decimal number.

In another aspect, the present invention provides a character data detection system using Adaboost learning, wherein step (c) comprises: (c1) grouping the candidate areas and extracting a representative candidate area from each group; (c2) extending the representative candidate region by a predetermined length; (c3) performing binarization on the extended representative candidate region; (c4) labeling the binarized representative candidate region; (c5) performing filtering using the size and proportion of the labeled area; (c6) detecting corner points in the region where labeling is performed, and selecting four vertices located at the outermost corners among the extracted corner points; And (c7) determining the final character area using the lengths of the left and right sides of the corresponding area with respect to the candidate area where the vertex is detected.

In addition, in the character data detection system using Adaboost learning, in the step (c1), the candidate areas are grouped by position and size, and are most similar to the average square position and average square size. And extracting a rectangle as a representative candidate region of the character.

In the character data detection system using adaboost learning, in the step (c2), the length and width of the representative candidate region are extended by the length k, and k is 1 / time of the representative candidate region. It is characterized by setting to 2.

In the character data detection system using Adaboost learning, in step (c3), binarization is performed using Otsu's adaptive method for the representative candidate areas. The threshold used for binarization is characterized by using a higher value than the average pixel value of the representative candidate region.

In addition, the present invention is a character data detection system using ada boost learning, in the step (c5), the size of the labeled area is out of a preset range, or the ratio of the aspect ratio of the labeled area in advance If outside the set range, it is characterized in that the removal from the representative candidate area.

In addition, in the character data detection system using ada boost learning, in the step (c6), for each pixel in the labeled region, the value of the pixel is a predetermined threshold and a predetermined threshold value of the peripheral pixel of the pixel; If the difference is greater than the value, the corresponding pixel is set as a candidate corner, and the brightness difference with 16 pixel values of a circle having a radius of 3 around the pixel P with respect to the pixel P of the candidate corner is calculated and the pixel is outside the predetermined threshold. If is greater than the predetermined threshold number, the candidate corner is identified as a corner, and 3 × 3 mask is covered around the corner to select the point with the most difference within the range, and the final corner is selected. It features.

In addition, in the character data detection system using adaboost learning, in the step (c7), the length of the left and right sides of the candidate region is determined as follows for the candidate region where the vertex is detected. If true by verifying the logical expression of, characterized in that the candidate area is detected as the final character area.

[Formula 2]

Moreover, this invention is characterized by the character data detection system using ada boost learning.

The present invention also relates to a computer-readable recording medium having recorded thereon a program for performing a method of detecting character data using Adaboost learning.

As described above, the character data detection method using Adaboost learning according to the present invention not only uses the MCT feature and the Adaboost method but also effectively filters the candidate area by the geometrical feature, thereby more accurately than the prior art. The effect of detecting them is obtained. Specifically, through experiments, it can be seen that the text areas can be extracted more robustly by 2.1% than the conventional methods from various input images.

In addition, according to the character data detection method using Adaboost learning according to the present invention, the effect that can be very useful in the real application fields related to multimedia and image processing, such as signage sign recognition of shops, license plate recognition of cars, etc. are obtained. .

 The method according to the present invention uses the MCT feature and the AdaBoost method, and can effectively detect the character regions because the candidate region is effectively filtered using geometric features.

BRIEF DESCRIPTION OF THE DRAWINGS The figure which shows the structure of the whole system for implementing this invention.
2 is a flowchart illustrating a method of detecting character data using adaboost learning according to an embodiment of the present invention.
3 is a diagram showing an example of MCT feature extraction according to an embodiment of the present invention.
4 is a table showing a pseudo-code of the ada boost method according to an embodiment of the present invention.
5 is a flowchart illustrating a method of detecting a final character area using geometric features according to an embodiment of the present invention.
6 is an exemplary image of grouping candidate rectangles according to an embodiment of the present invention.
7 is a flowchart illustrating a FAST method according to an embodiment of the present invention.
8 is an exemplary image of a training image including a license plate used for MCT feature extraction and Adaboost learning and a background image not including a license plate according to an experiment of the present invention. b) Example image of the image for learning.
9 is an exemplary image of a result of labeling candidate regions of a license plate extracted from a test image according to an experiment of the present invention.
10 is an exemplary image showing license plate areas of a vehicle finally detected except non-character areas through filtering of candidate areas according to an experiment of the present invention.
FIG. 11 is a graph showing a performance comparison result of accuracy in terms of a conventional method and a character area extraction method of the present invention according to an experiment of the present invention.

DETAILED DESCRIPTION Hereinafter, specific contents for carrying out the present invention will be described with reference to the drawings.

In addition, in describing this invention, the same code | symbol is attached | subjected and the repeated description is abbreviate | omitted.

First, examples of the configuration of the entire system for implementing the present invention will be described with reference to FIG.

As shown in FIG. 1, in the character data detection method using Adaboost learning according to the present invention, a computer terminal 20 receiving an image (or image) 10 and performing character data detection on the image (or image) Can be implemented as a program system. That is, the character data detection method may be configured as a program and installed and executed in the computer terminal 20. The program installed in the computer terminal 20 may operate like one program system 30.

On the other hand, in another embodiment, the character data detection method using Adaboost may be implemented by a program, which operates on a general-purpose computer, and may be implemented by an electronic circuit such as an ASIC (custom semiconductor). Alternatively, the present invention may be developed as a dedicated computer terminal 20 which processes only text data detection for an image. This is called the detection device 40. Other possible forms may be implemented.

On the other hand, the image 10 is composed of frames that are continuous in time. One frame has one image. Also, the image 10 may have one frame (or image). That is, the image 10 is also applicable to one image.

Next, a character data detection method using adaboost learning according to an embodiment of the present invention will be described with reference to FIG. 2.

In the present invention, candidate regions of a character are extracted from an input image by using a Modified Census Transform (MCT) feature [Non Patent Literature 8-9] and an Adaboost Learning Method [Non Patent Literature 10].

As shown in FIG. 2, first, preprocessing is performed on an input image (S10). In particular, Gaussian smoothing is performed to remove noise that may be partially included in the input image.

Next, MCT (Modified Census Transform) features are extracted from the preprocessed input image (S20).

The MCT feature is a local structure feature and is represented by binary information of 0 and 1. In other words, the MCT feature represents a relationship between a pixel at a specific position and pixels adjacent to the periphery. In general, the MCT feature is robust to changes in illumination because the MCT feature expresses the correlation between the pixel values located at a constant position in a predetermined pattern instead of using the pixel values themselves at a predetermined position. And because the calculation is relatively simple, the detection rate is relatively high and the execution time is fast in image processing applications.

In the present invention, the MCT feature Γ (x) using a 3x3 kernel may be expressed as in Equation 1.

[Equation 1]

는 커널 내에 위치한 화소들의 평균 화소 값을 나타낸다. n'은 커널의 중심과 인접한 화소들의 집합을 나타낸다. ζ()는 비교 함수로서 I(y)가 평균 화소 값 보다 크면 1을 출력하고, 그렇지 않으면 0을 출력한다. 그리고

는 10진수 변환 연산자로서 비교 함수 ζ()의 결과로 나오는 아홉 자리의 2진수 배열을 10진수로 변경해 준다. 따라서 본 발명에서 사용하는 MCT 특징은 0에서 511까지의 범위를 갖는다. 또한, MCT 특징은 입력영상과 같은 크기의 영상으로 생성된다.In Equation 1 above, I (x) represents a pixel value of x,

Denotes an average pixel value of pixels located in the kernel. n 'represents a set of pixels adjacent to the center of the kernel. ζ () is a comparison function and outputs 1 if I (y) is greater than the average pixel value, and 0 otherwise. And

Is a decimal conversion operator that converts the nine-digit binary array resulting from the comparison function ζ () into a decimal number. Therefore, the MCT features used in the present invention range from 0 to 511. In addition, the MCT feature is generated as an image having the same size as the input image.

3 graphically illustrates an example of an MCT transformation.

Next, by inputting the MCT feature extracted in the previous step to the character detection classifier generated by the Adaboost learner to detect candidate regions of the character region in the image (S30).

Adaboost learning is a method of boosting, which means combining a number of weak learners with poor classification performance to build a strong learner with good classification performance. Weak classifiers usually refer to classifiers with more than 50% accuracy, while strong classifiers are classifiers with very small classification errors.

In general, the Ada Boost method is widely used in the field of image processing and pattern recognition because it is relatively easy to implement and fast speed [Non-Patent Document 11]. This method also has the advantage that it can be used in combination with other types of learning methods to further improve its performance.

The Adabooth learning method generates a set from the original data according to the error rate of the learned hypothesis, and then uses it to learn a new hypothesis and construct a data set from the learned hypotheses. In addition, we study hypotheses and focus on data that is difficult to classify by reducing the weights of hypothetically predicted data and increasing the weights of incorrectly predicted data according to hypotheses. The AdaBoost method actually has less learning errors than the theoretical upper limit, and generalization errors also decrease with learning errors.

The table of Figure 4 shows the pseudo-code of the Ada Boost method.

In the table of FIG. 4, N denotes the number of samples, w _j denotes the weight of the j th sample, and K denotes the number of classifiers. c _k represents the k-th classifier, ε represents the error, and α _k represents the reliability of the classifier c _k . A distinctive feature of the AdaBoost method is that the classifiers are complementary. In other words, c _{k + 1} compensates for the weakness of c _k .

Next, only the final character areas are detected by verifying the candidate areas of the characters extracted using the Adaboost learning method in the previous step by using geometric features (S40). Detecting the final character area using geometric features is shown in FIG. 5.

As shown in FIG. 5, first, candidate regions are grouped and a representative candidate region is determined in each group (S41).

As shown in FIG. 6, quadrangles representing candidate areas of a character detected using the Adaboost method appear intensively in the vicinity of the actual character area. Therefore, in the present invention, grouping is performed first by the position and size of the rectangle, and then, in each group, the rectangle most similar to the average rectangle position and the average rectangle size is determined as the representative candidate area of the letter.

In other words, grouping means grouping candidate rectangles having similar positions and sizes. Rectangles representing candidate regions of characters extracted using the Adaboost learning method are concentrated in a number of squares near the actual character region. Therefore, all rectangles having positions and sizes within a predetermined range are grouped with respect to the positions and sizes of the rectangles. In FIG. 6, yellow squares represent candidate areas of letters, and squares displayed in blue color represent representative candidate areas of each group.

The character candidate area processed below is performed only for the representative candidate area.

Next, in consideration of the case where the detected character candidate area does not include all of the originally existing character areas, the horizontal and vertical lengths of the candidate area (or representative candidate area) are extended by a predetermined length k (S42). In the present invention, k is set to 1/2 of the height of the character candidate area.

Next, binarization is performed on the candidate regions (or representative candidate regions) of the extended character using an adaptive method proposed by Otsu (S43). In this case, the background part of the candidate area of the letter except for the letter generally includes a lighter color than the letter, so the threshold used for binarization is lower than the average pixel value of the letter candidate area. Use a high value.

Next, labeling is performed on candidate regions (or representative candidate regions) of the binarized character (S44). That is, a unique number is assigned (labeled) to each pixel of the binarized candidate region, and the adjacent pixel regions are labeled with the same object or the same unique number by analyzing the connectivity of each pixel. In particular, adjacent pixel regions connected to each other have the same unique number, and other adjacent region components have another unique number.

Then, a minimum rectangle including all of the areas labeled with one independent number (grouped and labeled pixels) is set as the labeled area.

The final purpose of performing such binarization and labeling is to obtain a character area (for example a license plate area of a car). In other words, the present invention does not search for the characters existing in the character area (number plate, etc.), but performs character recognition, but finds the character area (number plate area, etc.) containing the characters. Therefore, when the binarization is performed in the above step, as shown in FIG. When the labeling is performed, a labeled region is extracted with respect to the character region (eg, license plate region) indicated by white in FIG. 9. Then, the following operations are performed such as extracting corner points for the labeled character area (number plate area, etc.).

Next, filtering is performed using the size and ratio of the labeled area (S45). That is, if the size of the labeled region is too small or does not match the aspect ratio of the region, it is removed from the character candidate region. For example, license plates have a constant size and a constant aspect ratio. However, if the size or ratio is outside the predetermined range, it is not the license plate, so these candidate areas are excluded from the candidate.

Next, a corner point is detected in the area where labeling is performed, and four vertices located at the outermost part of the extracted corner points are selected (S46). In the present invention, FAST, a fast corner extraction method proposed by Edward Rosten of the University of Cambridge, UK, is used to extract corner points (Non-Patent Document 12).

FAST, as the name suggests, is a method of feature extraction that seeks speed. And the great thing about FAST is that even though FAST is designed to be optimized for speed, the quality also exceeds that of existing methods.

Basic processing of the FAST method may be expressed as shown in FIG. 7.

The FAST method consists of three steps as shown in FIG. The first step is to quickly select corner candidates. In this step, it is to check how much difference one point P has from surrounding points. In other words, if the difference is greater than a predetermined threshold value, the candidate is determined.

In the second step, the brightness difference is examined for all the points that become candidate corners. In other words, the candidate corner point P is judged by looking at 16 fire velocity values on a circle having a radius of 3 with respect to the point P. That is, by examining the difference in brightness with the points located around P, if there are more than k dark or light points, the point P is determined as a corner.

In the third step, a 3x3 mask is placed around the point P and the point with the most difference within the range is selected as the final corner. In other words, one problem of the FAST corner is that if a point P is recognized as a corner point, neighboring points adjacent to P are often detected as corner points. Therefore, FAST applies an additional post-processing step, the third step, to solve this problem.

Next, the length of the left side and the right side of the corresponding region is compared to the candidate region of the character where the vertex is detected as shown in Equation 2 to finally verify the actual character region (S47).

[Equation 2]

In Equation 2, Len _l represents the length of the left side of the rectangle representing the candidate area, and Len _r represents the length of the right side of the rectangle. And R _i denotes the i th candidate character area.

Len _l and Len _r are the lengths of the left or right sides of the rectangle, calculated from the vertices.

Next, the effect of this invention is demonstrated through experiment.

The computer used for the experiment in the present invention is composed of an Intel Core (TM) i7 2.93Ghz CPU and 8GB of main memory, using a Windows 7 operating system (OS). And the method of the present invention was implemented using Microsoft's Visual Studio and OpenCV open library. In this experiment, various types of test images including text areas were collected and used to compare and evaluate the performance of the method of the present invention.

8 shows an example of a learning image including a license plate and a background image without a license plate used for MCT feature extraction and AdaBoost learning in the experiment of the present invention. In the present invention, learning was performed using 307 license plate images and 2,736 background images.

9 shows an example of a result of labeling candidate regions of a license plate extracted from a test image. 9 shows the license plate areas of the vehicle finally detected except the non-character area by filtering the candidate areas. In FIG. 9, green squares represent candidate regions of the detected character region, and blue squares represent actual character regions selected through filtering. The red square represents the character area of the license plate finally extracted through image correction.

In order to quantitatively evaluate and evaluate the performance of the character region extraction method such as a license plate according to the present invention, the ratio of the extracted character regions not including the non-character region and the entire character regions present in the image is a percentage (%). The same accuracy measure as Equation 3 is used. In Equation 3, N _detected represents the number of character regions accurately detected using the method of the present invention, and N _existing represents the total number of character regions _existing in the test image.

[Equation 3]

FIG. 10 is a graph illustrating a performance comparison result of accuracy of the character region extraction method obtained through Equation 3. FIG. As can be seen in Figure 10 it can be seen that the method using the method of the present invention more accurately detects the character region. In other words, in the conventional method using texture information using high frequency elements, the binarization between the background and the text area is not properly performed, and thus, a lot of false detection of the text area occurs. However, the method according to the present invention uses the MCT feature and the Adaboost method, and since the candidate region is effectively filtered using geometric features, the character regions can be detected more accurately.

Recently, there is an increasing need for a study that analyzes input indoor and outdoor images and then strongly partitions only the text areas included in the images, such as license plates.

In the present invention, a new method for robustly detecting only text areas excluding background areas from various input images using MCT feature and AdaBoost method is proposed. To do this, first, candidate regions of characters were extracted using MCT features and Adaboost learning methods. Then, the candidate areas of the extracted text were filtered using geometric features to robustly detect only actual text areas except non-text areas. Experimental results show that the proposed method detects text areas more accurately than the conventional method.

As mentioned above, although the invention made by this inventor was demonstrated concretely according to the said Example, this invention is not limited to the said Example and can be variously changed in the range which does not deviate from the summary.

10: image 20: computer terminal
30: program system

Claims (10)

In the character data detection method using ada boost learning, detecting the character region in the input image,
(a) extracting a Modified Census Transform (MCT) feature from the input image;
(b) detecting candidate regions of a text area in the input image by inputting the extracted MCT feature to the text detection classifier using a text detection classifier generated by an Adaboost learning method; And,
and (c) filtering the candidate regions to detect a final character region using geometric features of a candidate region.
The method of claim 1,
MCT feature Γ (x) of the input image is extracted by the following [Equation 1] character data detection method using ada boost learning.
[Equation 1]

Where I (x) represents the pixel value of x,

Denotes the average pixel value of the pixels located in the kernel, n 'denotes the set of pixels adjacent to the center of the kernel, ζ () is a comparison function, and outputs 1 if I (y) is greater than the average pixel value; Is a function that outputs zero,

Is a decimal conversion operator that converts the nine-digit binary array resulting from the comparison function ζ () into a decimal number.
The method of claim 1, wherein step (c) comprises:
(c1) grouping the candidate regions and extracting a representative candidate region from each group;
(c2) extending the representative candidate region by a predetermined length;
(c3) performing binarization on the extended representative candidate region;
(c4) labeling the binarized representative candidate region;
(c5) performing filtering using the size and proportion of the labeled area;
(c6) detecting corner points in the region where labeling is performed, and selecting four vertices located at the outermost corners among the extracted corner points; And,
and (c7) determining the final character area using the lengths of the left and right sides of the corresponding area with respect to the candidate area where the vertex is detected.
The method of claim 3,
In step (c1), the candidate areas are grouped by position and size, and the Adaboost learning is characterized by extracting a rectangle most similar to an average rectangle position and an average rectangle size as a representative candidate area of a character. Character data detection method using a.
The method of claim 3,
In the step (c2), the length and width of the representative candidate region is extended by the length k, and k is set to 1/2 of the representative candidate region. .
The method of claim 3,
In the step (c3), binarization is performed on the representative candidate regions using Otsu's adaptive method, and the threshold used for binarization is an average of the representative candidate regions. Character data detection method using adaboost learning, characterized in that the higher value than the pixel value.
The method of claim 3,
In step (c5), if the size of the labeled area is out of a preset range or if the ratio of the aspect ratio of the labeled area is out of a preset range, the representative candidate area is removed. Character data detection method using boost learning.
The method of claim 3,
In the step (c6), for each pixel in the labeled area, if the value of the pixel has a difference larger than a predetermined threshold value from the surrounding pixel value of the pixel, the pixel is designated as a candidate corner, and With respect to the pixel P, a brightness difference with 16 pixel values of a circle having a radius 3 around the pixel P is calculated. A character data detection method using Adaboost learning, characterized in that the 3 × 3 mask is covered around the corner, and the point with the most difference within the range is selected as the final corner.
The method of claim 3,
In step (c7), if the vertex is detected and the length of the left and right sides of the candidate region is verified by verifying the following logical expression, the candidate region is detected as the final character region. Character data detection method using ada boost learning, characterized in that.
[Formula 2]

A non-transitory computer-readable recording medium having recorded thereon a program for performing the method of claim 1.

Images