CN1190755C - Colour-picture damage-free compression method based on perceptron - Google Patents

Abstract

A perceptron-based lossless compression method for color images relates to the field of image processing. The present invention is characterized in that it has the following processing steps: (1) the computer reads in the target image to be processed from an interface such as USB and stores it in the internal memory; (2) carries out existing lossless color space transformation to the image to be compressed; (3) adopts The two-dimensional weighted prediction model predicts the current pixel and obtains the prediction residual value; (4) according to the mapped prediction residual value, it is judged whether it is less than a given error limit, and if it exceeds the threshold, adaptive adjustment is required. (5) Adopt existing RICE entropy coding; (6) Output and save the compression result of the image. The complexity of the method provided by the present invention is lower than that of the traditional compression method, the processed image has a higher compression ratio, and while maintaining a higher compression effect, its compression time is significantly less than the latter, and it has a higher execution efficiency. speed.

Description

Color Image Lossless Compression Method Based on Perceptron

technical field

The invention relates to the field of image processing, and designs and implements a color image lossless compression method based on a neural network perceptron.

Background technique

With the wide application of multimedia technology in recent years, image lossless compression technology plays an important and irreplaceable role in some important image information storage and transmission applications, which can greatly reduce the impact on storage media capacity and transmission bandwidth requirements. Especially in the application system that is repeatedly stored, in order to maintain the original quality of the image, it is generally required to use lossless compression technology.

Since there is redundant information among image pixels, the purpose of image compression can be achieved by utilizing the correlation between image data and re-encoding to remove redundant information in image data. As for the color image, it can be regarded as a sequence of multi-component monochrome data images scanned or interpolated and arranged. (b) and (c) in Figure 1 are the autocorrelation function of pixels in each color plane and the cross-correlation function of pixels between each color plane in the standard image (a), respectively. Comparing Figure (b) and Figure (c), it can be seen that the correlation between adjacent pixels in the same monochromatic plane is higher than the correlation between corresponding pixels in different monochromatic planes. For Figure (a), the correlation between G plane and B plane pixels is higher than the correlation between R and G or R and B. According to this characteristic of color images, a color image compression scheme can be designed.

For the lossless compression of still images, the international organization ITU-T (International Telecommunications Union--International Telecommunication Union) was formerly known as CCITT (International Telephone and Telegraph Consultative Committee---International Telephone and Telegraph Advisory Committee) and ISO (International Organization for Standardization)/IEC (International Electrotechnical Commission) jointly formulated an international standard for lossless compression in 1992, and recommended the predictor and Huffman code based on DPCM (Differential Pulse Code Modulation---linear predictive coding) as the early version of the lossless compression coding algorithm JPEG standard. However, these traditional DPCMs or fixed-mode predictors that are simply improved cannot obtain completely irrelevant prediction data. The compression ratio of the algorithm is relatively low, especially the compression time is long, which is far from meeting the needs of current applications.

A lossless compression scheme for image data generally includes two parts: modeling and coding. The emergence of the arithmetic coding method can separate the two parts conceptually, and assign the data codewords after the model mapping from the probability. And researchers can concentrate on researching and designing various efficient prediction models, so that lossless compression can be greatly improved compared with the early JPEG standard algorithm in terms of compression ratio and compression efficiency. A new standard accepted by the JBIG/JPEG (Joint Photographic Experts Group) committee: JPEG-LS is a lossless/nearly lossless image compression algorithm that uses the LOCO-I (LOw COmplexity LOssless COmpression for Images) algorithm as its core content. The algorithm uses two-dimensional prediction, and adaptively corrects the predicted value by introducing context, and performs Golomb coding on the obtained residual signal. When the predictor finds that there are consecutive identical pixels, it switches from the normal mode to the run length encoding mode. This algorithm can reach more than 2.5 times the general image compression ratio, and the compression efficiency is high, so it has attracted widespread attention. But the context modeling in the JPEG-LS algorithm increases the complexity of the algorithm.

The present invention introduces the perceptron technology in the existing neural network, corrects the parameters of the two-dimensional weighted prediction model proposed by the present invention through the self-learning and self-adaptive ability of the perceptron, and simplifies the algorithm complexity of the prediction model part. In the entropy coding part, considering that the currently commonly used entropy coding methods mainly include Huffman, arithmetic coding, LZW, etc. These entropy encodings, while having efficient encoding results, are slow to perform. The present invention adopts the existing RICE entropy coding algorithm. The entropy encoding algorithm can calculate and encode in real time according to the sampling value during the scanning process of the data, without resorting to a code table stored in advance, which improves the speed of entropy encoding. In the encoding process, it also has an adaptive function to adapt to the change of code length. It has low complexity, fast coding speed and high coding efficiency.

Contents of the invention

The present invention designs and implements a lossless compression method for the entire color image by proposing a two-dimensional weighted prediction model (referred to as the prediction model). This method overcomes the shortcomings of slow execution speed, high algorithm complexity and low compression efficiency in previous image lossless compression algorithms. While taking into account the lower algorithm complexity, it achieves a fast and efficient image lossless compression effect, and the compression ratio is high. According to the analysis of the correlation between color components and pixels of color image, firstly, the existing lossless color space transformation is used to process the color space, and the correlation between color components is eliminated. In the two-dimensional weighted prediction model proposed by the present invention, the perceptron technology in the existing neural network is introduced. Using its own self-learning and self-adaptive features, it can make self-adaptive adjustment of the two-dimensional weighted forecasting model. Make the algorithm run with small prediction residuals. Further, according to the prediction residual mapping algorithm proposed by the present invention, the prediction residual value is mapped, and the dynamic range of the prediction residual value is reduced. It ensures that the algorithm has high execution efficiency under the premise of high compression ratio. After the prediction residual is obtained, entropy coding is required. The coding method in this paper adopts the RICE algorithm in the prior art. The entropy coding algorithm has low complexity, fast coding speed and high coding efficiency. During the execution of the entire compression algorithm, it is guaranteed that no quantization error will be introduced, and the original data will be completely restored during the decoding process, thereby realizing lossless image compression.

Technical thought feature of the present invention is:

1. For the correlation between color planes, the existing lossless color space transformation method is used to perform image preprocessing on the target image to remove the correlation of color space, so that the compression ratio can be effectively improved.

2. Assuming that the image data is in the order from left to right and from top to bottom, three color component data after space transformation are sequentially input. Since there is a strong correlation between adjacent pixels of the same component, the predicted value of the current pixel can be obtained through the prediction algorithm proposed by the present invention according to the weighted values of the four adjacent pixels above and to the left of the current pixel.

3. If this forecast value is very close to the current actual value, the difference between them is called the forecast residual. Find the prediction residual value, if the value is small enough, the encoder can encode it with a very short codeword.

4. In the actual algorithm, the prediction residual value is not directly encoded, but firstly undergoes a mapping process proposed by the present invention. Through the data mapping algorithm, the dynamic range of the prediction residual value is reduced, and the distribution range of the prediction residual signal used for coding is more concentrated, so that a smaller entropy can be obtained and a better coding effect can be obtained.

5. According to the result of the predicted residual value after mapping above, judge whether the predicted residual value is less than the given error limit 8. If the predicted residual value after mapping exceeds this threshold, it is necessary to modify the predicted weighting coefficient, that is, to realize the perception-based The adaptive adjustment subroutine of the device, if the mapped prediction residual does not exceed this threshold, RICE coding is performed on the mapped prediction residual value.

6. During the prediction and encoding process, a perceptron technique is used to adaptively correct the parameters of the two-dimensional weighted prediction model. The prediction residual value is monitored through the self-learning and adaptive capabilities of the perceptron, and if the value exceeds the set threshold, adaptive adjustment is made so that the prediction residual value remains in a small range during the entire encoding process In order to achieve the purpose of compression.

Refer to Fig. 3 and Fig. 4 for the technical solution of the present invention. This perceptron-based color image lossless compression method is to collect the target image to be processed by a digital camera or other digital instruments, and convert the optical signal of the target image into a digital signal image and input it to the computer for processing and transmission. . The computer processing mainly uses the existing USB interface software to compress the image on the basis of the existing color image lossless compression and the perceptron technology in the neural network. The processed result image is output to the buffer, which can be directly stored locally or remotely transmitted through a network storage device. The present invention is characterized in that it also comprises the following steps:

1) The computer reads in the target image to be processed from the USB interface and saves it in the memory;

2) preprocessing the image to be compressed, that is, carrying out existing lossless color space transformation;

3) Use a two-dimensional weighted prediction model to predict the current pixel, obtain the prediction residual value, and map the prediction residual value, which is characterized by:

The flow chart of this part of the process is shown in area A in FIG. 6 .

a) In order to describe the whole algorithm, it is assumed that the image adopts RGB color space and is preprocessed by color space transformation. Sequentially input the three transformed color components, that is, sequentially scan and store the data of the transformed three color components from left to right and from top to bottom. The length and width of the image are H and B respectively, and its pixels are represented by x _k (i, j). Here i=0, 1,..., B-1 represents the row where the pixel is located in a certain color plane, j=0, 1,..., H-1 represents the column, k=R, G, B, represents the pixel after transformation color plane.

公式为：Four pixels x _k (i, j-1), x _{k (i-1, j-1), x k ( i} -1, j), x _k ( i-1, j+1) weighted value to predict the current value, where i represents the row coordinates, j represents the column coordinates, and their distribution positions are shown in Figure 5, in which x ₁ , X ₂ , x ₃ , and x ₄ are representatives x _k (i, j-1), x _k (i-1, j-1), x _k (i-1, j), x _k (i-1, j+1) four pixels, X is the current pixels. Judging whether the predicted pixel is an edge point, when the predicted pixel is in the first column, the last column or the first row, special processing is performed, that is, the value of the pixel that does not exist outside the edge of the image is replaced by the default value; for the current Predicted value of pixel x _k (i, j)

The formula is:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, w _p (p=1, 2, 3, 4) is the prediction weight coefficient, and its initial value is set to w _p =0.25, from

And get the predicted value of the current pixel

b) According to the prediction residual value formula ②, the prediction residual value Δ is obtained:

$\mathrm{<span\; class="notranslate">\; \&Delta;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

If the prediction is guaranteed to be accurate during the entire encoding process, a small enough prediction residual value can always be obtained, which is beneficial to subsequent entropy encoding. The flow chart of this part of the process is shown in area C in FIG. 6 .

c) After the prediction residual is obtained, the prediction residual mapping method is as follows: assuming that each monochrome pixel is quantized by 8 bits, the distribution range of the prediction residual value should be [-255, +255]. Execute the following mapping algorithm, so that the distribution range of the prediction residual value becomes [0, +255] and the prediction residual value is represented by X. The entire process of mapping is shown in Figure 7.

i. Calculate the prediction residual value X according to the prediction model;

ii. When the predicted residual value X is less than -128, add 256 to it or when the predicted residual value X is greater than 127, subtract 256 to obtain a new predicted residual value X1;

iii. If the new prediction residual value X1 is less than zero, use it according to X2=-X1×2-1, otherwise use it according to X2=2×X1 to obtain the mapped prediction residual value X2;

iv. Output the predicted residual value X2 after mapping;

Through the data mapping algorithm, the dynamic range of the prediction residual value is reduced, and the distribution range of the prediction residual signal used for coding is more concentrated, so that smaller entropy can be obtained and better coding effect can be obtained;

4) According to the above mapped prediction residual results, judge whether the prediction residual is less than a given error limit 8, if the prediction residual value exceeds this threshold, it is necessary to modify the prediction weighting coefficient, that is, to realize the adaptive adjustment based on the perceptron subroutine, if the predicted residual after mapping does not exceed this threshold, proceed to step 5);

The subroutines for perceptron-based adaptive tuning are characterized as follows:

If only the above-mentioned fixed prediction model is used, it cannot be guaranteed that the predicted value will not shift during the entire coding process, so that the prediction residual used for coding will drift. In the algorithm proposed by the present invention, the double-layer perceptron in the neural network is used to monitor and adaptively adjust the drift of the prediction residual value in real time.

The training process of the perceptron and the modification method of the prediction weighting coefficient are as follows:

The network structure of a two-layer perceptron is shown in Figure 2, which consists of an input layer and an output layer. Each processing unit of the input layer is interconnected with each processing unit of the output layer, and weights are attached to each connection, and these weights can be adjusted by learning rules. Its learning process is the process of changing the connection strength between neurons.

The output value can be represented by the following formula:

$\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Among them, _f (.) is the activation function, w ₁ , w _{2 . 1} , x ₂ ... x _n is the input of the processing unit p, and y is the output value;

The training of the perceptron is carried out with a set consisting of a set of samples. During the training period, these samples are repeatedly sent to the input layer of the perceptron, and the output of the perceptron is adjusted to achieve the ideal output by adjusting the connection weight. The training process is as follows:

a) Initialize, assign an initial value to each w _p , and assign w _p (t) a random value in the interval (-1, +1); where w _p (t) represents the connection weight on the pth input at time t value, 1≤p≤n, w ₀ (t) is the threshold at time t;

b) Correction of connection weights, input a sample X=(x ₁ , x ₂ ... x _n ) and its expected output d;

i. Calculate the actual network output y(t) according to the formula ③;

ii. Calculate the error between the actual output y(t) of the network and the expected output d:

                                                     

iii. Modify the connection weight wp between each input and output:

w _p (t+1)＝w _p (t)+Δw _p (t) ⑤

Δw _p (t) = α×x _p ×d(t) (p=1, 2, 3, 4) ⑥

Among them, α is the gain, usually 0≤α≤1, which is used to control the correction speed; Δw _p (t) is the correction value of the connection weight w _p at time t, and x _p is the pth pixel value currently input;

c) Repeat step ② for each sample until the error is less than a given error limit;

After learning, the neural network memorizes the learning sample pattern in the form of connection weights. When the neural network inputs a certain pattern, the neural network will calculate the output value y according to formula ③, so the whole learning and memory process is based on the actual output and the desired output. The error adjustment parameter w _p between formula ① and formula ③ can be seen, here the activation function: f(x)=x, x _p in formula ③ is the p-th pixel value of the current input, and w _p is the predicted The weighting coefficient, y is the prediction value; the perceptron monitors the prediction residual value in real time, if this value exceeds the preset threshold during the encoding process, a feedback signal is generated to the perceptron, and by relearning and adjusting w _p , a A new set of prediction weighting coefficients, in the actual image coding process, the formula is:

Δw _p =α/128×x _p ×Δ (p=1,2,3,4) ⑦

Among them, Δw _p is the correction value of the connection weight w _p at time t, α=0.001 is the empirical value, and Δ is the prediction residual value of the previous pixel; in this way, the predictor uses the revised prediction weighting coefficient to obtain a more accurate predicted value;

Since the perceptron has the ability of learning memory and adaptively adjusting connection weights, it is obviously possible to use this model to adaptively adjust the predictive weighting coefficients in the image compression algorithm to obtain a higher compression ratio than the fixed predictive weighting coefficients. Adaptively adjust the prediction weighting coefficients through the difference between the actual output and the expected output, which is the learning process of the perceptron.

The perceptron monitors the prediction residual value in real time. If the value exceeds the preset threshold during the encoding process, a feedback signal is generated to the perceptron. By relearning and adjusting w _p , a new set of prediction weighting coefficients is generated. Use formulas ⑤ and ⑥ to modify the prediction weighting coefficients. The flow chart of this part of the processing process is shown in area B in Figure 6.

It can be seen that in the adaptive adjustment algorithm, the perceptron model adopts a mode of learning while working, without prior training. The image is scanned once, and the prediction and encoding are done at the same time. Has high execution efficiency;

5) Using the RICE entropy coding algorithm to code the mapped prediction residual;

6) Output and save the code stream of image compression.

When performing preprocessing, that is, performing an existing lossless color space transformation on the image, the principle is as follows:

For multi-component (color) images, through a certain lossless color space transformation, the correlation between color components can be removed to a certain extent. JPEG-LS, JPEG2000 and other color image compression methods have carefully designed color space transformation methods. The space transformation in the present invention is carried out according to formula ①. Among them, A1, A2, and A3 are the three color components before transformation, respectively, and the three color component values obtained after transformation of A1', A2', and A3'.

The complexity of the method proposed in the present invention is lower than that of the traditional compression method, the processed image has a higher compression ratio, and while maintaining a higher compression effect, its compression time is significantly less than the latter, and it has a higher execution efficiency. speed. In the entropy coding part, the RICE entropy coding algorithm is used. During the scanning process of the data, the entropy coding algorithm can calculate and code in real time according to the sampling value, without resorting to the code table stored in advance, which improves the speed of entropy coding. . In the encoding process, it also has an adaptive function to adapt to changes in different code lengths. It has low complexity, fast coding speed and high coding efficiency.

Description of drawings:

Figure 1 is a color image correlation analysis diagram:

(a) standard image; (b) autocorrelation function of pixels in the same component plane; (c) cross-correlation function of pixels between different component planes;

Fig. 2 is a structure diagram of a double-layer perceptron;

Fig. 3 is a block diagram of a fast color image lossless compression system based on a perceptron;

1. Digital camera, 2. USB interface, 3. Computer processor, 4. Output cache, 5. Lossless image compression encoding, 6. Compressed result file, 7. Hard disk, 8. Network storage device, 9. Decoding, 10. Display;

Fig. 4 is method flowchart of the present invention;

Fig. 5 is a position diagram of four pixels and the current pixel in the predictor;

Fig. 6 is a flow chart of the prediction module and adaptive adjustment to the current pixel;

Fig. 7 is the mapping module flowchart of prediction residual value;

Fig. 8 is a decoding flowchart;

Fig. 9 is the original image of the embodiment, and its RGB components;

Fig. 10 is three color components after preprocessing, a picture is A1 ' component, b picture is A2 ' component, and c picture is A3 ' component;

Fig. 11 is the subroutine flowchart of image compression algorithm;

Fig. 12 is a program flow chart of the present invention.

Detailed ways

The digital video camera and USB interface in Figure 3 are commercially available, mainly to complete the collection of various images that need to be compressed, and convert the optical signal of the image to be processed into a digital signal image input to the computer, which is convenient for computer processing, transmission and other operations. Computer processing mainly reads in the collected images through the existing USB interface software, and the processed image files are output to the buffer memory of the computer, which is convenient for local storage (stored in the hard disk); data transmission and storage are carried out through network storage devices , processing; or directly decode, wherein the flow chart of decoding is shown in Figure 8, and then displayed. The lossless compression coding of images is realized by software. The whole process of image compression will be described in detail below in conjunction with specific examples, see FIG. 12 . The steps of the whole process can be seen in the technical idea features of the aforementioned invention, and the implementation process can be seen in Figure 12 and Figure 11:

First, the target image is collected through a digital camera, and the digital camera captures a frame of image when the car enters and exits during the garage management process of the community, and then transmits it to the hard disk of the computer through the USB interface, thus completing the collection of the monitoring image of the embodiment process.

After the image is transferred to the computer in the specific implementation, the following procedures are completed in the computer:

The first step: Fig. 9 is the original image of the embodiment, that is, the image (24-color true color, jpg format) and its RGB components to be compressed by the user, and the image size is (142×107). Perform lossless color space transformation on the image to obtain the result image after removing the correlation between color components (this part is independent of the main program). The processed results of the three components of A1, A2, and A3 are respectively shown in Figure 10 in Figure a, Figure B, and Figure c;

Step 2: Open the resulting image from the previous step.

Step 3: Create a compressed file to store the compressed data.

The fourth step: extracting the image information of the image transformed by the lossless color space.

The 5th step: execute image compression algorithm subroutine, concrete steps are shown in the image compression algorithm flow chart of Fig. 11:

1) Read in a line of data and judge whether it is a boundary point. If it is a boundary point, then perform special processing on it, that is, replace the value of a pixel that does not exist outside the edge of the image with a default value; if it is not a boundary point, assign a value to the predictor , the initial value of the prediction weight coefficient array of the predictor is: W[1]=0.375; W[2]=0.375; W[3]=0.125; W[4]=0.125.

2) Using a two-dimensional weighted prediction model to predict the current pixel, obtain the prediction residual value, and map the prediction residual value;

a) Using the feature that there is a strong correlation between adjacent pixels of the same component, according to the four adjacent pixels on the upper left of the current pixel (assuming that the image data is in order from left to right and from top to bottom, sequentially Input three color component data after lossless color space transformation), and the predicted value of the current pixel can be obtained through a certain prediction algorithm.

b) Calculating the difference between the predicted value and the actual value is called the prediction residual. When the value is small enough, the encoder encodes it with a very short codeword;

c) Perform residual mapping processing on the obtained prediction residual. Through the data mapping algorithm, the dynamic range of the prediction residual value is reduced, and the distribution range of the prediction residual signal used for coding is more concentrated, so that a smaller entropy can be obtained and a better coding effect can be obtained.

3) According to the result of the above-mapped prediction residual value, judge whether the prediction residual value is less than a given error limit 8. If the mapped prediction residual value exceeds this threshold, it is necessary to modify the prediction weighting coefficient, that is, call the perception-based The adaptive adjustment subroutine of the device, if the predicted residual error after mapping does not exceed this threshold, then proceed to step 4).

In this process, the method proposed by the present invention using the existing perceptron technology to adaptively correct the parameters of the two-dimensional weighted prediction model is adopted. The prediction residual value is monitored through the self-learning and self-adaptive ability of the perceptron. If the value exceeds the set threshold, adaptive adjustment and mapping are performed so that the prediction residual value remains at a very low level during the entire coding process. In a small range, so as to achieve the purpose of compression. The entire flow chart of adaptively adjusting the parameters of the two-dimensional weighted prediction model is shown in FIG. 6 . Here, the number of neurons n is 1, the number of input nodes m is 4, and the gain α is 0.00001.

4) Existing RICE entropy coding is performed on the mapped prediction residual.

5) Determine whether it is the end of the line, if the line has ended, then save the current line; otherwise, determine whether it is a boundary point again, and repeat the above steps;

6) Judging whether it is at the end of the column, if it is at the end of the column, then end the subroutine; otherwise, repeat (1) to (7) steps;

Step 6: Print: "Mission completed!", "CompressedCounter=number of encoded bytes", "See compressed data in file compress.dat"

Step 7: Close the file to be compressed by the user and the compressed file.

Step 8: Get the compressed file and store it in the local hard disk, thus completing the storage of the monitoring records of the garage.

In order to test the performance of the algorithm proposed by the present invention, it is compared with the traditional Huffman algorithm, LZW algorithm, FELICS algorithm in the literature, especially the new lossless compression standard JPEG-LS. The images used for testing are all 24-bit color images in RGB space (8bits per pixel component). Includes two landscape images, one aerial image, and one computer-generated image. In order to make the results more explanatory, two standard color images peppers and lena are also selected for comparison experiments. The results of Huffman algorithm, LZW algorithm and JPEG-LS algorithm are obtained by running the execution program downloaded from the webpage, and the results of FELICS algorithm are obtained from the literature Barequet R, Feder M.SICLIC: A simple inter-color lossless image coder.In Proc .of data compression conference, Edited by J.A.Storer, Snowbird Utah, USA, March, 1999, pp.500-510. The whole experiment is completed with C language on a PC with 300MHz Pentium II and 64M memory.

Table 1 shows the comparative compression results, which are represented by the number of bits per pixel (=8bit/compression ratio). It can be seen from Table 1 that the compression result of the method proposed by the present invention is obviously better than the LZW algorithm and the traditional Huffman algorithm, and is also better than the result of the FELICS algorithm. Although the compression result of JPEG-LS is better than the result of the algorithm of the present invention, the results of the two are quite close, especially for two landscape images (smooth natural images), the result of the present invention is only 3% worse than the algorithm of JPEG-LS about.

From the compression time shown in Table 2, the execution speed of the algorithm of the present invention is obviously faster than the traditional Huffman and LZW algorithms, and 13-17% faster than the low-complexity and high-efficiency JPEG-LS algorithm. Comparing from the structure of the algorithm, this algorithm and JPEG-LS are all equivalent in prediction and encoding, but the context model (context modeling) is added in the JPEG-LS algorithm, and its complexity is far greater than the perceptron prediction used in the present invention model, although it improves the compression ratio, execution speed suffers.

Table 1: Comparison of compression results image Width Height Compression result (bits/pixel) Huffman ZW FELICS JPEG-LS this invention landscape 1 2048*1536 2.92 5.01 * 2.13 2.20 landscape 2 2056*2400 2.27 4.20 * 1.83 1.89 aviation 2200*1196 2.65 3.92 * 1.90 2.22 computer synthesis 1636*1070 2.47 3.08 * 2.21 2.55 Peppers 512*512 4.67 5.84 5.14 3.88 4.28 Lena 512*512 5.12 6.78 4.84 4.52 4.76

Table 2: Comparison of compression times image Width Height Compression time (s) Huffman ZW JPEG-LS this invention landscape 1 2048*1536 6.88 9.87 6.21 5.10 landscape 2 2056*2400 9.35 13.32 8.68 7.27 aviation 2200*1196 6.83 7.90 6.62 5.55 computer synthesis 1636*1070 5.24 8.68 4.46 3.89 Peppers 512*512 2.18 3.12 0.60 0.51 Lena 512*512 2.87 3.89 1.05 0.90

Claims (2)

1, a kind of coloured image lossless compression method based on perceptron, be to finish by Digital Video or other digitizers to gather pending target image, and the optical signalling of target image is converted to the digital signal image, computing machine reads in image by prior USB, interface such as infrared, in computer processor, image is handled, result images after the processing outputs to buffer, can directly store or carry out remote transmission, it is characterized in that it also comprises the steps: by the network storage equipment in this locality

1) computing machine is kept at the internal memory after USB interface is read in pending target image;

2) treat compressed image and carry out pre-service, promptly it is carried out existing harmless color notation conversion space;

3) adopt a kind of two-dimentional weighted prediction model that current pixel is predicted, ask for the prediction residual value,, it is characterized by this prediction residual value mapping:

A) order is imported three components behind the color notation conversion space, and from left to right promptly, sequential scanning and deposit data through three color components after the conversion successively from the top down, image length and width are respectively H and B, its pixel x _k(i j) represents;

By current pixel x _k(i, j) four the pixel x in upper left side _k(i, j-1), x _k(i-1, j-1), x _k(i-1, j), x _k(i-1, prediction weighting coefficient j+1) is predicted currency, and wherein i represents row-coordinate, and j represents the row coordinate;

Judge whether predicted pixel is marginal point, when predicted pixel is in first row, last row or first row, then carry out special processing, promptly substitute with default value for the value that does not have pixel outside the image border; To current pixel x _k(i, predicted value j)

Computing formula be:

$\stackrel{\&OverBar;}{x_k(i,j)}=w_1{x}_k(i,j-1)+w_2{x}_k(i-1,j-1)+w_3{x}_k(i-1,j)+w_4{x}_k(i-1,j+1)---\left(1\right)$

W wherein _p(p=1,2,3,4) are the prediction weighting coefficient, and setting its initial value is w _p=0.25, thus the predicted value of current pixel obtained

B) 2., obtain prediction residual value Δ by prediction residual value formula:

$\mathrm{\&Delta;}=x_k(i,j)-\stackrel{\&OverBar;}{x_k(i,j)}---\left(2\right)$

C) after obtaining prediction residual be: suppose that each include monochrome pixels is to quantize through 8bit to its method of carrying out the prediction residual mapping, then the distribution range of prediction residual value should be [255, + 255], mapping algorithm below carrying out, make its distribution range become [0, + 255] the prediction residual value is represented with X

I. calculate prediction residual value X according to forecast model;

Ii. when prediction residual value X less than-128 the time, with its add 256 or when prediction residual value X greater than 127 the time, it is deducted obtains a new prediction residual value X1 after 256;

If iii. new prediction residual value X1 less than zero with it according to X2=-X1 * 2-1, otherwise the prediction residual value X2 after it is obtained shining upon according to X2=2 * X1;

Iv. the prediction residual value X2 after output is shone upon;

4) according to the prediction residual value result after the above mapping, judge that whether prediction residual is less than the given limits of error 8, if the prediction residual value after the mapping exceeds this threshold value, just need to revise the prediction weighting coefficient, promptly realize adjusting subroutine based on the self-adaptation of perceptron, if the prediction residual after the mapping does not surpass this threshold value, then carry out step 5);

The feature of adjusting subroutine based on the self-adaptation of perceptron is as follows:

Use double-deck perceptron in the neural network of the prior art to come real time monitoring and adjust the drift of prediction residual adaptively, output valve can be represented by following formula:

$y=f\left(\underset{p=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_p{x}_p\right)---\left(3\right)$

Wherein, f (.) is an excitation function, w ₁, w ₂... w _nFor from input processing unit p (p=1,2 ..., n) to output processing unit the connection weights, w ₀Be threshold value, x ₁, x ₂... x _nBe the input of processing unit p, y is an output valve;

The training of perceptron adopts the set of being made up of one group of sample to carry out, and at training period, these samples is repeated to deliver to the input layer of perceptron, connects weights and makes the output of perceptron reach desirable to export by adjusting, and its training process is as follows:

A) each wp initialize is given in initialization, composes and gives wp (t) with (1 ,+1) interval random value; Here wp (t) expression t is p the connection weights of importing constantly, and 1≤p≤n, w0 (t) are t threshold value constantly;

B) a sample X=(x is imported in the correction of connection weights ₁, x ₂... x _n) and its desired output d;

I. the 3. actual output of computational grid y (t) by formula;

Ii. the error between the actual output of computational grid y (t) and the desired output d:

d(t)＝d-y(t) ④

Iii. revise the connection weights wp between each input and output:

w _p(t+1)＝w _p(t)+Δw _p(t) ⑤

Δw _p(t)＝α×x _p×d(t) (p＝1，2，3，4) ⑥

Wherein, α is gain, and common 0≤α≤1 is used to control erection rate; Δ w _p(t) connect weight w constantly for t _pModified value, x _pP pixel value for current input;

C) 2., to each sample repeating step until error less than the given limits of error;

Neural network after study finishes memorizes the form of learning sample pattern with connection weight, when neural network is imported a certain pattern, 3. neural network will calculate output valve y with formula, and therefore whole learning and memory process is exactly the error adjustment parameter w according to reality output and between wishing to export _p, relatively formula 1. with formula 3. as can be seen, excitation function: f (x)=x here, the formula x in 3. _pBe p pixel value of current input, w _pBe the prediction weighting coefficient, y is a predicted value; Perceptron real time monitoring prediction residual value if this value has exceeded pre-set threshold in cataloged procedure, then produces one and feeds back signal to perceptron, by relearning and adjust w _p, producing one group of new prediction weighting coefficient, in the real image cataloged procedure, formula is:

Δw _p＝α/128×x _p×Δ (p＝1，2，3，4) ⑦

Wherein, Δ w _pFor t connects weight w constantly _pModified value, α=0.001 is an empirical value, Δ is the prediction residual value of a last pixel; Like this, fallout predictor uses corrected prediction weighting coefficient, obtains predicted value more accurately;

5) adopt RICE entropy coding algorithm that the prediction residual after shining upon is encoded;

6) code stream of output and preservation compression of images.

2, the coloured image lossless compression method based on perceptron according to claim 1 is characterized in that treating the harmless color notation conversion space formula that compressed image carries out being adopted in the pre-service and is:

Wherein, A1, A2, A3 are respectively three color components before the conversion, three color component value that obtain after A1 ', A2 ', the A3 ' conversion.

Images