CN107507250A - A kind of complexion tongue color image color correction method based on convolutional neural networks - Google Patents

Abstract

A method for color correction of face color and tongue color image based on convolutional neural network relates to a digital image processing method. The algorithm mainly includes offline part and online part. The offline part consists of training data collection, color correction convolutional neural network framework construction and training, and the online part is image color correction and color correction effect evaluation. CNN imitates the human cognitive process, that is, layer-by-layer abstraction from local features to global features. Applying convolutional neural networks to color correction can achieve a more ideal color reproduction effect. The present invention uses a deep convolutional neural network method to perform color correction on the facial complexion and tongue color collected in a stable optical environment, and reproduce the color information that the target object actually presents in the same optical environment.

Description

A method for color correction of complexion and tongue color images based on convolutional neural network

technical field

The invention relates to a digital image processing method, in particular to a method for correcting complexion and tongue color images based on a convolutional neural network.

Background technique

The collection and reproduction of true colors is of great value in fields such as medicine and art. Image color information is an important basis for some professional image analysis. The color presented on the surface of an object is closely related to various links such as light source characteristics, lighting conditions, acquisition equipment, display equipment, and printing equipment. Color correction is a key technology for color reproduction and consistent color presentation. At present, color correction has been applied in many image processing fields such as medical images, mural images and license images. It is of great significance to study the color correction technology that can truly reflect the color of the observed object itself.

The color distortion in the camera-acquired image compared to the actual scene image. Therefore, many methods for image color correction have been proposed one after another, such as polynomial regression, color correction based on partial least squares regression, and neural networks. The polynomial regression color correction method requires fewer training samples and has low computational complexity, but its regression accuracy is closely related to the selection of training samples and polynomials, and its extrapolation ability is poor. The partial least squares regression color correction method can better solve problems such as multiple correlations between independent variables and a relatively small number of samples, but the accuracy is still difficult to meet the needs of actual medical applications. The training of traditional neural network-based color correction methods is limited by the number of network layers, initialization parameter selection and parallel mode, so the generalization performance of the network generally has over-fitting phenomenon.

In recent years, deep learning has been widely used, among which Convolutional Neural Network (CNN) is a typical deep feedforward network. CNN imitates the human cognitive process, that is, layer-by-layer abstraction from local features to global features. Applying convolutional neural networks to color correction can achieve a more ideal color reproduction effect.

Contents of the invention

The purpose of the present invention is to perform color correction on the facial complexion and tongue color collected in a stable optical environment by using a deep convolutional neural network method, and to reproduce the color information that the target object actually presents in the same optical environment.

The present invention is realized by adopting the following technical means:

A color correction method for complexion and tongue color images based on convolutional neural networks, the overall flow chart is shown in Figure 1; the algorithm mainly includes an offline part and an online part. The offline part consists of training data collection, color correction convolutional neural network framework construction and training, and the online part is image color correction and color correction effect evaluation.

The specific content of the offline part is as follows:

(1) Training data collection

The image acquisition of the present invention adopts a closed environment—artificial dark box, to avoid the influence of stray light from the outside, and artificial light source is used for lighting to ensure the quality and stability of tongue image acquisition; the relative positions of the light source and imaging equipment are fixed, so as to achieve The consistency and standardization of the environment; the artificial light source (D65) is used to simulate natural light under the dark box condition, effectively ensuring the stability of the light source condition.

Unlike conventional methods, this method uses ColorChecker Digital SG as a color card for color correction. ColorChecker Digital SG has 140 color blocks, which has a wider color gamut than the traditional ColorChecker Classic. At the same time, in the training samples, we increase the color patch samples that are similar to the skin color and tongue color, which is conducive to improving the accuracy of color correction. Under closed environmental conditions, take pictures of ColorChecker Digital SG standard color cards. By changing the shooting angle of the color card, adjusting the distance between the color card and the light source, adjusting the distance between the color card and the camera, etc., the color card images are captured, and these images are used to generate training data for the color correction model of the convolutional neural network. Process the captured image, crop and intercept each color block, each color block needs to be set in a fixed size format as a training sample, use the standard value of each color block of the color card to generate an RGB image as the label of the training data, training samples and labels One to one correspondence.

(2) Construction and training of color correction convolutional neural network network framework

In the present invention, the neural network is used to fit the relationship between the real color and the image color obtained by photographing, and the learning content is relatively simple, so the network is designed as a shallow deep neural network with 5 layers. The present invention adopts three different layer structures, as shown in Figure 2, which are input layer, nonlinear transformation layer and output layer respectively. The input layer is composed of a convolutional layer and a rectified linear unit (ReLU); the nonlinear transformation layer consists of a 3-layer network, each layer consists of a convolutional layer and a ReLU activation function, and the convolutional layer and activation There is a batch normalization in the middle of the function; the output layer is composed of a convolutional layer.

In the training, the present invention uses the stochastic gradient descent algorithm with mini-batch to iterate and update the convolution kernel state W and bias B, each time a micro-batch data set (mini-batch) operation is performed, and the stochastic gradient descent algorithm is used Find the global optimal solution.

In the image processing process of CNN, convolutional layers need to be connected through convolutional filters. The definition of convolutional filters is expressed as W×H×C×D, where C represents the number of channels of the filtered image; W, H represents the width and height of the filtering range respectively; D represents the type of convolution filter.

The input layer of the network consists of a convolutional layer and a ReLU activation function. The input layer feature extraction formula is expressed as follows:

F ₁ (X ₁ )=max(0, W ₁ *X ₁ +B ₁ ) (1)

where _X1 is the feature map entering the input layer. W ₁ and B ₁ represent the convolution filter and bias of the input layer respectively. The size of W ₁ is 3×3×3×64, which represents 64 different convolution filters, and the kernel size of each convolution is 3 ×3×3, F ₁ (X ₁ ) is the feature map obtained by the input layer. The input image is a 3×40×40 feature map, which means that the feature map is a 3-channel color map, and the width w and height h are both 40. The calculation formulas of width w ₁ and height h ₁ of the output feature map through the convolutional layer are shown in formula (2) and formula (3), kernel is the kernel size of convolution; stride is the step size of convolution kernel, when the value When it is 1, the overlapping image blocks are extracted, and the effect is better; pad is the number of zero-filled pixels on the edge. In the present invention, the value of kernel is set to 3, the value of stride is 1, and the value of pad is 1. Therefore, after the input image passes through the 64 convolution kernels of the input layer 3×3, a 64×40×40 feature map will be generated; then, the feature map will pass through the modified linear unit ReLU. The representation of ReLu is max(0,X), which can extract useful feature maps. The final output is still a 64×40×40 feature map.

In the nonlinear mapping process of the nonlinear transformation layer, the convolutional layer, batch normalization and ReLU functions are located in the second, third and fourth layers. The formulas of each stage of the nonlinear transformation layer are expressed as follows:

F _i (X _i )=max(0, W _i *F _i-1 (X _i-1 )+B _i ){i=2, 3, 4} (4)

In formula (4), i represents the i-th layer, and X _i is the output of the i-1th layer, that is, F _i-1 (X _i-1 ). W _i and B _i respectively represent the convolution filter and bias of the nonlinear transformation stage, where the size of the convolution filter W ₁ is 3×3×3×64, and the second, third, and fourth layers of convolutional layers W The size of _i is 64×3×3×64, and the size of each convolution kernel is 64×3×3. The 64×40×40 feature map output by the input layer is input into the second convolutional layer, and a 64×40×40 feature map will be generated after passing through 64 convolution kernels 3×3. Then, the 64×40×40 feature maps go into batch normalization. Batch normalization is between the convolutional layer and the ReLU activation function, which solves the problems of slow convergence and gradient explosion during neural network training. At the same time, batch normalization speeds up the training speed of the network and improves the model accuracy. Finally, the feature maps are modified with linear units to improve the non-linearity of the features. The second layer network outputs a 64×40×40 feature map, and then passes through the third and fourth layers with the same structure as the second layer, and finally obtains a 64×40×40 feature map.

In the output reconstruction process of the output layer, the feature map is input to the output layer which contains only one convolutional layer. The formula for the output reconstruction is expressed as follows:

F ₅ (X ₅ )＝W ₅ *F ₄ (X ₄ )+B ₅ (5)

In the formula, X ₅ is the output of the fourth layer. W ₅ and B ₅ represent the convolution filter and offset of the feature reconstruction layer respectively. The size of W ₅ is 3×3×64×3, and the feature reconstruction layer has 3 convolution filters, which are equivalent to the function of the mean filter , the kernel size of each convolution is 3×3×64, which can realize the function of the average feature map, F ₄ (X ₄ ) is the feature map generated by the nonlinear transformation layer, that is, X ₅ ; the feature map output by the nonlinear transformation layer After 3 convolution kernels 3×3, a 3×40×40 feature map will be generated.

The collected data set is trained by the network, and the model of each round of training is obtained after more than 50 iterations, and the model is finally saved to a file.

The online part, the specific content is as follows:

(1) Image color correction

The color-distorted image is color-corrected by using the trained model to obtain the corrected image. In the darkroom of the present invention, the color card, face and tongue images are taken, and the obtained photos are distorted compared with the actual color, and the color correction method based on the convolutional neural network is used to correct the color of the distorted image. First, read the pixel points of the image to be corrected and save it as an image matrix, and then read the MAT format file obtained from the training to obtain the color correction model. Input the image matrix into the network model, perform color correction on the image in the R, G, and B channels respectively, and output the corrected image.

(2) Evaluation of color correction effect

In order to verify the validity of the color correction model, it is necessary to evaluate the color correction effect. The evaluation of color correction is a very complex issue, involving different disciplines such as colorimetry, physiology, and psychology. Commonly used evaluation methods include objective evaluation and subjective evaluation.

According to the theory of colorimetry, the evaluation criteria of color reproduction include reflection spectrum matching, color appearance matching and tri-stimulus matching, etc., which are all objective standards. Tristimulus value matching is to make the tristimulus value of the object displayed by the computer the same as the tristimulus value of the corresponding actual object color. The quality of color reproduction is good if the tristimulus value of the captured image is the same as that of the corresponding actual object or the color difference is within the allowable range. Tristimulus value matching is the most common and practical color reproduction criterion. The present invention adopts CIE1976L*a*b* as an evaluation indicator.

Subjective evaluation means that someone directly evaluates the visual experience of a given stimulus. In the present invention, the observer of the color correction subjective evaluation compares the real food with the corrected image, and then evaluates whether the method restores the true color.

Description of drawings:

Figure 1 is a flow chart of color correction method based on convolutional neural network;

Figure 2 Architecture diagram of color correction convolutional neural network model;

Figure 3 Comparison of face and tongue before and after correction.

detailed description

According to the above description, the following describes the specific implementation process of the present invention.

The offline part is divided into 2 steps:

Step 1: Training Data Acquisition

Image acquisition is the basis for color correction work. When the acquisition equipment and lighting conditions change, how to ensure that the acquired image has constant color characteristics is a key issue in image acquisition. It involves many issues such as the design of the image acquisition device, the selection of the lighting source, the selection of the color space (Color Space), the establishment of the mathematical model of the system color characteristics, and so on. Therefore, the standardization of color correction image acquisition environment and method is an important basis for color correction. Generally, a dark room or a dark box is the most ideal shooting environment. The self-developed dark box adopted by the present invention can avoid the interference of external stray light and keep the light source environment stable.

Step 1.1: Normalize Sample Collection

After a large number of tests, in order to ensure the quality of image acquisition and avoid the influence of the external environment, the present invention proposes the following conditions for the color correction acquisition environment.

1) Use a closed collection environment to avoid stray light entering the shooting environment and strong light entering the lens;

2) The experimental light source is D65 light source, simulating natural light;

3) D65 light source stabilization time is 10 minutes. Turn on the light source, and collect the sample image after the light source is stable;

4) Fix the positions of the light source and the camera, and set the position of the color card within 30-35 cm from the camera;

5) Correctly set the camera parameters. For Canon EOS1200D, set the white balance to auto, the aperture to F10, the ISO to 3200, and the shutter time to 1/160 second;

Step 1.2: Sample Collection

The sample collection process is carried out according to the set conditions. The present invention uses the Canon EOS1200D with configured parameters to take pictures of the ColorCheck Digital SG color card in the dark box, and obtains a large number of color card photos by changing the position of the color card and the shooting angle of the camera. In order to increase the robustness of color correction, the face and tongue are photographed under the same lighting environment, and the obtained face and tongue images can be applied to the verification of the correction effect of the convolutional neural network color correction model. The ColorCheckDigital SG color card has optical standard values, so the obtained color card photos, after processing, correspond to the corresponding standard images one-to-one, and can be used as training samples for the color correction network.

Step 1.3: Sample Preprocessing

The ColorCheck Digital SG color card photos taken by the camera cannot be directly used as training samples for the convolutional neural network, and each color block needs to be segmented. The present invention divides each color block of the color card, and the size of each color block is 180×180 pixels. The LAB value of each color block of the color card can be obtained by consulting the official ColorCheck Digital SG optical data. The invention converts the LAB value of each color block into the RGB value under the D65 light source environment, and uses the RGB value of the color block to generate an RGB image. Similarly, the present invention sets the size of the standard color block RGB image to 180×180, and the size of the color block image cut from the photo is the same as that of the standard color block image, which is convenient for subsequent calculations. The RGB image of the standard color card is used as the training label of the intercepted color patch image, and the intercepted color patch image is in one-to-one correspondence with the label name, which can avoid large errors caused by corresponding errors.

In order to improve the generalization ability and robustness of the convolutional neural network color correction model, the present invention splices 140 color blocks cut from the same color card image according to the order of the color blocks in the ColorCheck Digital SG color card to generate a splicing The images are put into the training data together with the single color block pictures, and the standard color block photos are stitched together in the same order and put into the label data set. In this way, the types of training data samples are diverse, which is conducive to improving the generalization ability of the model.

In order to increase sample diversity and expand the sample library, the present invention performs data enhancement on the training data set by flipping up and down, rotating 90 degrees to the left, rotating 90 degrees to the right, converting 180 degrees to the right, rotating 90 degrees to the left and up and down Flip, rotate 90 degrees to the right and flip up and down, convert 180 degrees to the right and flip up and down seven transformations, expand the training data to 8 times the original size, and then use the sliding window to divide the training samples into 40*40 small images. After a series of expansions, the number of training data sets reached 96,000. Color correction model training is essentially an input-to-output mapping, which can learn a large number of mapping relationships between inputs and outputs, and finally generate a correction model. In order to verify the accuracy of the verification model, a test set with a size of 1280 images is prepared in a similar manner. The training set and test set are sent to the convolutional neural network for training.

Step 2: Building and training the color correction convolutional neural network framework

In the present invention, no pooling layer and fully connected layer are added, and the entire network is divided into an input layer, a nonlinear transformation layer, and an output layer. The input layer is composed of a convolutional layer and a ReLU activation function; the nonlinear transformation layer is composed of a 3-layer network, each layer is composed of a convolutional layer and a ReLU activation function, and there is a batch between the convolutional layer and the activation function Normalization; the output layer is composed of a convolutional layer.

(1) The input layer of the network contains a convolutional layer and a ReLU activation function. The input layer feature extraction formula is expressed as follows:

F ₁ (X ₁ )=max(0, W ₁ *X ₁ +B ₁ ) (6)

where _X1 is the feature map entering the input layer. W ₁ and B ₁ represent the convolution filter and bias of the input layer respectively. The size of W ₁ is 3×3×3×64, which represents 64 different convolution filters, and the kernel size of each convolution is 3 ×3×3, F ₁ (X ₁ ) is the feature map obtained by the input layer.

In the nonlinear mapping process of the nonlinear transformation layer, the convolutional layer, batch normalization and ReLU functions are located in the second, third and fourth layers. The formulas of each stage of the nonlinear transformation layer are expressed as follows:

F _i (X _i )=max(0,W _i *F _i-1 (X _i )+B _i ){i=2,3,4} (7)

In the formula, i represents the i-th layer, and X _i is the output of the i-1th layer. W _i and B _i respectively represent the convolution filter and bias of the nonlinear transformation stage, where the size of the convolution filter W ₁ is 3×3×3×64, and the second, third, and fourth layers of convolutional layers W The size of _i is 64×3×3×64, and the size of each convolution kernel is 64×3×3.

In the output reconstruction process of the output layer, the feature map is input to the output layer which contains only one convolutional layer. The formula for the output reconstruction is expressed as follows:

F ₅ (X ₅ )＝W ₅ *F ₄ (X ₄ )+B ₅ (8)

In the formula, X ₅ is the output of the fourth layer. W ₅ and B ₅ represent the convolution filter and offset of the feature reconstruction layer respectively. The size of W ₅ is 3×3×64×3, and the feature reconstruction layer has 3 convolution filters, which are equivalent to the function of the mean filter , the kernel size of each convolution is 3×3×64, which can realize the function of the average feature map, and F ₄ (X ₄ ) is the feature map generated by the nonlinear transformation layer.

In the process of model training, the input image size is 40×40 feature map, in the first convolutional layer, after 64 convolution kernels 3×3, a 64×40×40 feature map will be generated; in the second In the first convolutional layer, the input size is 64×40×40 feature map, after passing through 64 convolution kernels 3×3, a 64×40×40 feature map will be generated; in the third convolutional layer, the input size is It is a 64×40×40 feature map, and after 64 convolution kernels 3×3, a 64×40×40 feature map will be generated; in the fourth convolutional layer, the input size is 64×40×40. Figure, after 64 convolution kernels 3×3, a 64×40×40 feature map will be generated; finally input to the convolution layer of the output layer, after 3 convolution kernels 3×3, a 3×40×40 feature map will be generated feature map of .

(2) In the nonlinear transformation layer, batch normalization is added between the convolution layer and the excitation layer. The main idea of batch normalization is to whiten the input data, speed up network convergence, reduce data redundancy and feature correlation, and actually make the data 0 mean and unit variance through linear transformation. Batch normalization solves the untrainable problems such as slow convergence and gradient explosion encountered during neural network training. At the same time, batch normalization speeds up the training speed of the network and improves the accuracy of the model.

(3) In the activation function of each layer, the present invention uses a modified linear unit, and the formula is shown in formula (9). x represents the result of the feature map after the convolution kernel, when x<0, f(x)=0; if x>0, f(x)=x. During forward propagation, the calculation speed is accelerated. During backpropagation, when x>0, the gradient is 1, thus alleviating the gradient dispersion problem. Therefore, the convergence speed of the obtained stochastic gradient descent algorithm is greatly improved compared with sigmoid and tanh.

f(x)=max(0,x) (9)

(4) During network training, the present invention adopts the stochastic gradient descent algorithm with mini-batch for training. When the sample size is relatively large or the number of iterations is high, the traditional gradient descent algorithm has a slow operation speed, and the stochastic gradient descent algorithm overcomes these shortcoming. The traditional training brings in all samples each time, and the stochastic gradient descent algorithm with mini-batch is brought into the micro-batch data set (mini-batch) for operation, and uses random descent plus to find the global optimal solution. The learning rate is a necessary parameter of the stochastic gradient descent algorithm learning method, which determines the speed of weight update. If it is set too large, the cost function will oscillate, and the result will exceed the optimal value. If it is too small, the convergence speed will be too slow. Small learning rate, such as 0.001 ± 0.01 to keep the system stable. The momentum parameter and the weight decay factor can improve the training adaptability, the momentum parameter is usually [0.9,1.0], and the weight decay factor is usually 0.0005±0.0002. Through experimental observation, the present invention sets the learning rate to 10 ^-4 , the momentum parameter to 0.9, and the weight decay factor to 0.0005. After many experiments, the present invention sets the batch size to 128.

After the design and adjustment of the above four main steps, the convolutional neural network color correction model is built, and the training sample data collected in step 1 is put into the network for training. After the training is completed, the model after 50 rounds of training is obtained. Matlab's MAT format saves.

The online part is divided into 2 steps:

Step 1: Image Color Correction

Photograph the color card, human face and tongue in the dark box, and obtain the color card, human face and tongue photos under the light source environment. Through comparison, it can be found that there are different degrees of color distortion in photos and actual objects. The present invention uses a convolutional neural network-based color correction model to perform color correction on color-distorted images. First, the RGB value of the image to be corrected is read pixel by pixel and saved as an image matrix. Then, read the color correction network from the MAT format of Matlab to obtain a color correction model based on convolutional neural network. The image matrix is input into the network, and the three convolution kernels of the input layer read the matrix of the three channels of the image R, G, and B. Afterwards, the model performs color correction on the R, G, and B components of the image, respectively. After the three-channel image matrix is color-corrected by the network, the three-channel image matrix is output. The three image matrices are finally merged into an RGB image, that is, a color-corrected image is obtained.

Step 2: Color correction effect evaluation

The model is used to correct the color of the color card, face and tongue images, and the corrected images are obtained. In order to verify the effectiveness of the color correction model, the corrected images need to be evaluated.

The evaluation of color correction involves fields such as colorimetry and psychology, and is a complex issue. It is usually divided into two categories: subjective evaluation and objective evaluation. Subjective evaluation methods often allow the observer to observe and compare the corrected image and the photographed object. Observers measure how well the colors are reproduced by comparing the colors of the corrected image with the real colors. This method is intuitive and effective, and it is the main method for color correction quality evaluation.

Objective evaluation is carried out under certain conditions with a set of color scales. Color correction works better if the corrected color is close to the original color stop value. Generally, if the color reproduction tristimulus value and the color difference of the color scale are within the allowable range, it is considered that the corrected image and the real object have the same visual effect. In CIELAB space, it is generally believed that ΔE<3 is the standard for true color reproduction, and ΔE<6 is the standard for accurate and acceptable color correction.

In order to verify the effectiveness of the color correction method based on the convolutional neural network, the present invention performs color correction on the color card, face and tongue images; in order to evaluate the performance of the color correction model, the present invention respectively gives subjective and objective color Evaluation results.

Step 1.1: Subjective Evaluation

As shown in Figure 3, this experiment gives a comparison of the face and tongue before and after color correction. Canon cameras originally took photos of people with brownish and dark problems. After color correction, the color correction method based on the convolutional neural network can restore the true color of the face and tongue very well.

Step 1.2: Objective Evaluation

For the objective color evaluation results, for the standard color card image, use the self-developed software of the present invention to cut out the color block image of the color card image, and read the RGB value in each color block, convert the RGB value into a LAB value, and use the standard error Value CIE1976L*a*b* As an evaluation index, the error is shown in formula (10).

In the above formula, ΔL ^* , Δa ^* and Δb ^* are the difference between the LAB value of the color patch image and the LAB value of the standard color patch, respectively. Using the above method, test the chromatic aberration of the sample before and after correction. Calculated, the sample color error before correction Average is 14.21, color corrected for error The average is 3.70. Therefore, the color correction method proposed by the present invention has achieved good correction results.

Claims (2)

1. A complexion and tongue color image color correction method based on convolutional neural network, comprising offline part and online part, characterized in that: offline part is composed of training data collection, color correction convolutional neural network framework and training, online Some include image color correction; The specific content of the offline part is as follows: (1) Training data collection The acquisition adopts artificial light source under dark box conditions to simulate natural light, effectively ensuring the stability of light source conditions; Process the captured image, crop and intercept each color block, each color block needs to be set in a fixed size format as a training sample, use the standard value of each color block of the color card to generate an RGB image as the label of the training data, training samples and labels one-to-one correspondence; (2) Construction and training of color correction convolutional neural network network framework The network is designed as a shallow deep neural network, and the number of network layers is 5 layers; they are input layer, nonlinear transformation layer, and output layer; the input layer is composed of a convolutional layer and a modified linear unit ReLU; the nonlinear transformation layer consists of 3-layer network, each layer consists of a convolutional layer and ReLU activation function, there is a batch normalization between the convolutional layer and the activation function; the output layer is composed of a convolutional layer; In training, use the stochastic gradient descent algorithm with mini-batch to iterate and update the convolution kernel state W and bias B, perform micro-batch data set operations each time, and use the stochastic gradient descent algorithm to find the global optimal solution; In the image processing process of CNN, convolutional layers need to be connected through convolutional filters. The definition of convolutional filters is expressed as W×H×C×D, where C represents the number of channels of the filtered image; W, H represents the width and height of the filtering range respectively; D represents the type of convolution filter; The input layer of the network consists of a convolutional layer and a ReLU activation function. The input layer feature extraction formula is expressed as follows: F ₁ (X ₁ )=max(0, W ₁ *X ₁ +B ₁ ) (1) where _X1 is the feature map entering the input layer. W ₁ and B ₁ represent the convolution filter and bias of the input layer respectively. The size of W ₁ is 3×3×3×64, which represents 64 different convolution filters, and the kernel size of each convolution is 3 ×3×3, F ₁ (X ₁ ) is the feature map obtained by the input layer. The input image is a 3×40×40 feature map, which means that the feature map is a 3-channel color map, and the width w and height h are both 40. The calculation formulas of width w ₁ and height h ₁ of the output feature map through the convolutional layer are shown in formula (2) and formula (3), kernel is the kernel size of convolution; stride is the step size of convolution kernel, when the value When it is 1, the overlapping image blocks are extracted, and the effect is better; pad is the number of zero-filled pixels on the edge. In the present invention, the value of kernel is set to 3, the value of stride is 1, and the value of pad is 1. Therefore, after the input image passes through the 64 convolution kernels of the input layer 3×3, a 64×40×40 feature map will be generated; then, the feature map will pass through the modified linear unit ReLU. The representation of ReLu is max(0,X), which extracts useful feature maps. The final output result is still a feature map of 64×40×40; <mrow><msub><mi>w</mi><mn>1</mn></msub><mo>=</mo><mfrac><mrow><mi>w</mi><mo>-</mo><mi>p</mi><mi>a</mi><mi>d</mi><mo>*</mo><mn>2</mn><mo>+</mo><mi>ker</mi><mi></mi><mi>n</mi><mi>e</mi><mi>l</mi></mrow><mrow><mi>s</mi><mi>t</mi><mi>r</mi><mi>i</mi><mi>d</mi><mi>e</mi></mrow></mfrac><mo>+</mo><mn>1</mn><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow></mrow> <mrow><msub><mi>h</mi><mn>1</mn></msub><mo>=</mo><mfrac><mrow><mi>h</mi><mo>-</mo><mi>p</mi><mi>a</mi><mi>d</mi><mo>*</mo><mn>2</mn><mo>+</mo><mi>ker</mi><mi></mi><mi>n</mi><mi>e</mi><mi>l</mi></mrow><mrow><mi>s</mi><mi>t</mi><mi>r</mi><mi>i</mi><mi>d</mi><mi>e</mi></mrow></mfrac><mo>+</mo><mn>1</mn><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>3</mn><mo>)</mo></mrow></mrow> In the nonlinear mapping process of the nonlinear transformation layer, the convolutional layer, batch normalization, and ReLU functions are located in the second, third, and fourth layers; the formulas of each stage of the nonlinear transformation layer are expressed as follows: F _i (X _i )=max(o, W _i *F _i-1 ( _xi-1 )+R _i ){i=2, 3, 4} In formula (4), i represents the i-th layer, Xi _i is the output of the i-1th layer, that is, F _i-1 (X _i-1 ); W _i and B _i represent the convolution filter in the nonlinear transformation stage and bias, where the size of the convolutional filter W ₁ is 3×3×3×64, the size of the 2nd, 3rd, and 4th convolutional layer W _i is 64×3×3×64, and each volume The size of the product kernel is 64×3×3; the feature map of 64×40×40 output by the input layer is input into the second convolution layer, and after 64 convolution kernels 3×3, 64×40× 40 feature map; then, the 64×40×40 feature map enters batch normalization; batch normalization is between the convolutional layer and the ReLU activation function, which solves the problem of slow convergence and gradient explosion during neural network training. The training situation; at the same time, batch normalization speeds up the training speed of the network and improves the model accuracy; finally, the feature map is modified by the linear unit to improve the nonlinearity of the feature; the second layer of the network outputs a 64×40×40 feature After the graph, the third and fourth layers with the same structure as the second layer are passed through, and finally a 64×40×40 feature map is obtained; In the output reconstruction process of the output layer, the feature map is input to the output layer containing only one convolutional layer; the formula for output reconstruction is expressed as follows: F ₅ (X ₅ )＝W ₅ *F ₄ (X ₄ )+B ₅ (5) In the formula, X ₅ is the output of the fourth layer; W ₅ and B ₅ represent the convolution filter and bias of the feature reconstruction layer respectively, the size of W ₅ is 3×3×64×3, and there are 3 feature reconstruction layers The convolution filter is equivalent to the function of the mean filter. The kernel size of each convolution is 3×3×64, which can realize the function of the average feature map. F ₄ (X ₄ ) is the feature map generated by the nonlinear transformation layer. That is, X ₅ ; the feature map output by the nonlinear transformation layer will produce a 3×40×40 feature map after passing through 3 convolution kernels of 3×3; The collected data set is trained by the network, and the model of each round of training is obtained after more than 50 iterations, and the model is finally saved to the file; The online part, the specific content is as follows: Use the model obtained by training to correct the color distortion image to obtain the corrected image; take color cards, face and tongue images in the dark room, and the obtained photos are distorted compared with the actual color, and color correction based on convolutional neural network is used The method performs color correction on the distorted image; first reads the pixel points of the image to be corrected and saves it as an image matrix, and then reads the MAT format file obtained from training to obtain the color correction model; the image matrix is input into the network model, respectively in R, G, The three channels of B perform color correction on the image and output the corrected image. 2. The method according to claim 1, characterized in that: Use ColorChecker Digital SG as the color card for color correction; take pictures of the ColorChecker Digital SG standard color card under closed environmental conditions; by including changing the shooting angle of the color card, adjusting the distance between the color card and the light source, adjusting the color card and camera The color card images are obtained by shooting in distance mode, and these images are used to generate the training data of the convolutional neural network color correction model.