A method for generating defect sample images

Abstract

The invention discloses a defect sample image generation method. The defect sample image is generated by using a model. The training process of the model is as follows: a normal image is input into a neural network A to output a feature matrix Fa; a label is input into a neural network B to output a feature matrix Fb; a defect image is input into a neural network C to output a feature matrix Fc; a matrix obtained by using the feature matrices Fb, Fc and Fa is input into a neural network D, and the neural network D outputs a feature matrix Fd; the matrix Fd is then passed through a maximum pooling layer to obtain a weight matrix, and the weight matrix acts on the feature matrix Fd to generate a defect sample image; the feature matrix Fd is input into a neural network E, and the neural network E outputs a defect information matrix; a loss value is calculated; when the loss value reaches stability, the model parameters are saved; the method aims to improve the accuracy and diversity of generating defect sample images, and solves the problem that the generated defect sample images have a single form and high repeatability.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a method for generating defect sample images. Background Art

[0002] The probability of defects such as flaws appearing in industrial sites is low, and the number of defective images collected by cameras is small. When performing defect identification and analysis, it is difficult to identify and analyze defects well due to the small number of defective samples. For example, when using deep learning models for defect detection, a large number of defective sample images need to be obtained to form training data in order to train a defect recognition model with high accuracy. Therefore, in order to obtain a sufficient number of defective sample images, it is necessary to artificially generate sample images containing defects through image generation methods. At the same time, in order to improve the effectiveness of defective sample images, the artificially generated defective sample images need to have a high degree of correlation and similarity with real, possible defects.

[0003] At present, the commonly used defect sample image generation methods include: data enhancement and generative adversarial model (GAN), among which data enhancement will increase the noise of image samples and cause the deviation of the real distribution of data; while GAN training is difficult, it is easy to ignore the diversity of image samples, resulting in low quality of generated image samples. Patent document CN202110978527.1 provides a method and device for generating training images containing defects, obtains a mask image of a defective image, selects an injectable area of ​​the defective image in a defect-free image according to the mask image, and then fuses the defective image in the selected area to generate a training image containing defects, but the generated defects have a high similarity with the original defects, and the defect location is limited, which is not conducive to sample diversity. Patent document CN202311628239.9 proposes an image generation method, a model training method, a device and a storage medium, which generates a large number of converted defective images through a small number of real defective images and annotated defective images, so as to alleviate the problem of insufficient defect data and difficulty in covering all defective scenes in defect detection. However, this method consumes a large amount of resources, is costly, and is difficult to apply in practice.

[0004] It can be seen that generating high-quality defect sample image data is difficult and has low accuracy, and an efficient and accurate defect image generation method is urgently needed. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a defect sample image generating method; this method aims to improve the accuracy and diversity of generating defect sample images, and in generating defect sample images, the location and severity of defects can be arbitrarily designed; it solves the problem of single form and high repeatability of generating defect sample images.

[0006] The technical solution is as follows:

[0007] A defect sample image generation method, using a trained deep learning model to generate a defect sample image, the deep learning model includes a neural network A, a neural network B, a neural network C, a neural network D and a neural network E;

[0008] The dimensions of the output layers of neural network A, neural network C, and neural network D are the same;

[0009] The number of channels in the output layer of neural network B and neural network C is the same;

[0010] The dimensions of the output layers of neural network B and neural network E are the same;

[0011] The training steps of the deep learning model are as follows:

[0012] 1) Collect multiple images of the object to be tested in the same inspection scene, including various defect images and normal images;

[0013] Defect images of different defect categories are stored separately to form N defect image sets; N represents the number of defect categories;

[0014] Label each defect image, wherein the label includes the location coordinates, size information, type and severity level of the defect;

[0015] For each label, in a set of defect images with the same type, find defect images with the same and / or similar severity levels as those in the label and mark them as associated images;

[0016] A single normal image, a single label, and the associated images corresponding to the label are recorded as a group of training samples; different labels and different normal images are randomly combined to form multiple groups of training samples;

[0017] 2) Input a set of training samples into the deep learning model and perform the following processing:

[0018] ① The normal image is input into the neural network A, and the neural network A is used to extract the features of the normal image and output the feature matrix Fa corresponding to the normal image;

[0019] The label is input into a neural network B, and the neural network B is used to extract label features and output a feature matrix Fb corresponding to the label;

[0020] The defect image is input into a neural network C, and the neural network C is used to extract features of the defect image and output a feature matrix Fc corresponding to the defect image;

[0021] Apply the feature matrix Fb to the feature matrix Fc to obtain the feature matrix Fc';

[0022] The feature matrix Fc' is horizontally spliced ​​with the feature matrix Fa and the matrix after the horizontal splicing is input into the neural network D, the neural network D is used to extract the features of the matrix and output the feature matrix Fd;

[0023] The feature matrix Fd is then passed through the maximum pooling layer to obtain a weight matrix, and the weight matrix is ​​consistent with the feature matrix Fb in size; the weight matrix is ​​applied to the feature matrix Fd to obtain the feature matrix Fd';

[0024] The value of each element in the feature matrix Fd' is recorded as the gray value of the pixel point to generate a defect sample image;

[0025] ② Input the feature matrix Fd into the neural network E, which is used to extract the features of the feature matrix Fd, perform identification, and output a defect information matrix with the same dimension as the label;

[0026] The loss value is calculated using the similarity between the generated defect sample image and the normal image, and the similarity between the defect information matrix and the label;

[0027] 3) Determine whether the loss value has reached stability:

[0028] If not, switch to the next set of training samples and continue to step 2) to iterate the model parameters;

[0029] If yes, the deep learning model training is completed and the model parameters are saved.

[0030] Furthermore, the method of generating defect sample images using the trained deep learning model is as follows:

[0031] Setting label I, which includes: location coordinates, size information, type and severity of the defect to be generated;

[0032] In a set of defect images consistent with the defect type in label I, search for one or more defect images that are consistent with and / or similar to the severity level in the label;

[0033] Input any normal image, label I, and one or more defective images found into the trained deep learning model, and perform the processing of step ① to generate defective sample images;

[0034] The generated defect sample image complies with the information set in label I.

[0035] Preferably, in label I, the severity level of the defect can be set in the range of: the minimum level to the maximum level of the defect levels corresponding to the N types of defects.

[0036] Preferably, in step 1), the severity of the defect is evaluated as follows:

[0037] Perform the following processing on each defect image set to obtain the defect levels corresponding to N types of defects:

[0038] ① Calculate the grayscale mean G' of the defect area in a single defect image in the current defect image set;

[0039] Find the image area corresponding to the defect area in M ​​normal images and calculate the grayscale mean G of the found image area i , i=1,2...M; take M G i The median or mean is recorded as G'';

[0040] Calculate the difference between G'' and G';

[0041] ② Perform step ① on each defect image in the defect image set to obtain multiple difference values;

[0042] ③ Set the severity level of the defect according to the size of the difference.

[0043] Furthermore, in step ③, the severity level of the defect is set according to the size of the difference, including the following two methods:

[0044] Method 1: Sort multiple differences from small to large, and divide each S difference into a level;

[0045] Method 2: Perform K nearest neighbor clustering on multiple differences to obtain K clustering results, and divide K severity levels according to the clustering results.

[0046] Preferably, when there are multiple defect images input into the neural network C, a weight value is set in advance for each defect image;

[0047] Each defect image corresponds to a feature matrix, and the corresponding elements in multiple feature matrices are weighted and added to obtain the feature matrix Fc.

[0048] Preferably, the weights are designed in the following way:

[0049] The weight of defect images that are consistent with the severity level in the label is greater than the weight of defect images that are similar to the severity level in the label.

[0050] Preferably, the network models selected by neural network A, neural network B, neural network C, neural network D and neural network E are not completely the same.

[0051] Preferably, the neural network A is a convolutional neural network with more than 13 convolutional layers;

[0052] Neural network B uses a convolutional neural network with more than 3 convolutional layers;

[0053] Neural network C uses a convolutional neural network with more than 13 convolutional layers;

[0054] Neural network D uses a convolutional neural network with more than 5 convolutional layers;

[0055] The neural network E uses a convolutional neural network with more than 13 convolutional layers.

[0056] This method has the following characteristics:

[0057] This method uses multiple networks to respectively identify the feature rectangles corresponding to the normal image, label matrix, and defect image, and then reshapes the features through rectangle fusion to parse and generate defect sample images that meet the label information. The defect area in the generated image is integrated into the background area, and there is no fault at the edge, which is closer to the real defect feature. It can effectively improve the accuracy of generating defect sample images. In addition, in the defect sample images generated by this method, the location and severity of the defects can be arbitrarily designed; it solves the problem of single form and high repeatability of generating defect sample images. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the model training process in a specific implementation method. DETAILED DESCRIPTION

[0059] The technical solution of the present invention is described in detail below in conjunction with embodiments.

[0060] A defect sample image generation method uses a trained deep learning model to generate defect sample images, such as Figure 1 As shown, the deep learning model includes neural network A, neural network B, neural network C, neural network D and neural network E;

[0061] Among them, the dimensions of the output layers of neural network A, neural network C and neural network D are the same;

[0062] The number of channels in the output layer of neural network B and neural network C is the same;

[0063] The dimensions of the output layers of neural network B and neural network E are the same;

[0064] In the specific implementation, there is no special restriction on the types of each neural network, and the existing convolutional neural network models such as ResNet, VGG, LeNet, etc. can be used;

[0065] In order to improve the accuracy of the model, it is preferably implemented as follows: the network models selected by neural network A, neural network B, neural network C, neural network D and neural network E are not exactly the same; for example, neural network A and neural network B use ResNet; neural network C, neural network D and neural network E use ‌VGG.

[0066] In this embodiment, the training steps of the deep learning model are as follows:

[0067] 1) Collect multiple images of the object to be tested in the same inspection scene, including various defect images and normal images;

[0068] Defect images of different defect categories are stored separately to form N defect image sets; N represents the number of defect categories;

[0069] Label each defect image, wherein the label includes the location coordinates, size information, type and severity level of the defect;

[0070] For each label, in a set of defect images with the same type, find defect images with the same and / or similar severity levels as those in the label and mark them as associated images;

[0071] For example, the defect categories include scratches, pits, and dirt spots. The defect levels range from light to heavy, with a total of 5 levels: 1~5;

[0072] In specific implementation, the size of the label is N rows × 6 columns;

[0073] Label , where the first column represents the abscissa of the upper left corner of the minimum circumscribed rectangle of the defect, the second column represents the ordinate, the third column represents the width of the upper left corner of the minimum circumscribed rectangle of the defect, the fourth column represents the height of the minimum circumscribed rectangle of the defect, the fifth column represents the type of defect, and the sixth column represents the severity of the defect;

[0074] Specifically: the first line shows the information of scratch defects; the second line shows the information of pit defects;

[0075] All 0s in the third row indicate no defects. In specific implementation, fill in the label according to the type and number of defects in the image. When there is only one defect, only one row of the label has content, and the other rows are all 0s.

[0076] Furthermore, the associated images are labeled as follows:

[0077] In the defect image set of the scratch category, find defect images that are consistent with and / or similar to severity level 2 (such as levels 1 and 3) and mark them as associated images;

[0078] In the defect image set of the pit category, find defect images that are consistent with and / or similar to severity level 3 (such as levels 2 and 4) and mark them as associated images;

[0079] In a specific implementation, in order to improve the recognition rate of the deep learning model for the severity of defects and make the result of generating defect sample images more accurate, it is preferably implemented as follows: in the defect image set of the scratch category, defect images (multiple images) with consistent and similar severity levels are simultaneously searched and marked as associated images;

[0080] A single normal image, a single label, and the associated images corresponding to the label are recorded as a group of training samples; different labels and different normal images are randomly combined to form multiple groups of training samples;

[0081] 2) Input a set of training samples into the deep learning model and perform the following processing:

[0082] ① The normal image is input into the neural network A, and the neural network A is used to extract the features of the normal image and output the feature matrix Fa corresponding to the normal image;

[0083] Specifically, the normal image passes through the convolution layer, pooling layer, fully connected layer, and output layer of the neural network A in sequence, and outputs a feature matrix Fa of size C×W×H, where C represents the number of channels, and W and H represent the width and height of the feature map corresponding to the normal image, respectively.

[0084] Among them, in order to improve the accuracy, neural network A uses a convolutional neural network with more than 13 convolutional layers;

[0085] The label is input into a neural network B, and the neural network B is used to extract label features and output a feature matrix Fb corresponding to the label;

[0086] Specifically, the labels pass through the convolution layer, pooling layer, fully connected layer, and output layer of the neural network B in turn to obtain a feature matrix Fb (weight matrix) of size C×1×1, where C represents the number of channels. For each channel, there is a one-to-one correspondence with a weight value, which represents the influence of each channel on feature extraction.

[0087] Since the amount of data in the label is small, the model architecture of neural network B is relatively simple, and the number of convolutional layers greater than 3 can meet the recognition requirements.

[0088] The defect image is input into a neural network C, and the neural network C is used to extract features of the defect image and output a feature matrix Fc corresponding to the defect image;

[0089] Specifically, the defect image passes through the convolution layer, pooling layer, fully connected layer, and output layer of the neural network C in sequence to obtain a feature matrix Fc of size C×W×H.

[0090] Since the amount of defect image data is large, the number of convolutional layers of the neural network C is greater than 13 to meet the recognition requirements.

[0091] More specifically,

[0092] When there are multiple defect images (related images) input to the neural network C, a weight value is set for each defect image in advance;

[0093] In the neural network C, each defect image corresponds to a feature matrix (C×W×H). In specific implementation, the corresponding elements in multiple feature matrices are weighted and added to finally obtain a feature matrix Fc (C×W×H);

[0094] In more detail, the weights are designed as follows:

[0095] The weight of defect images that are consistent with the severity level in the label is greater than the weight of defect images that are similar to the severity level in the label.

[0096] For example, the feature matrix corresponding to the defect image of scratch and severity level 1 is , the weight is pre-set to 0.8 (the severity level of the image is consistent with the severity level in the label);

[0097] The feature matrix corresponding to the defect image of scratch and severity level 2 is , the weight is pre-set to 0.2 (the severity level of the image is similar to the severity level in the label);

[0098] The final feature matrix ;

[0099] The feature matrix Fb (C×1×1) is applied to the feature matrix Fc (C×W×H) to obtain the feature matrix Fc', that is, each channel of the feature matrix Fc is multiplied by its own weight. When the weight is large, the value of the channel feature map will increase accordingly, and the impact on the final output will also become greater; when the weight is small, the value of the channel feature map will be smaller, and the impact on the final output will also become smaller.

[0100] Concatenate the feature matrix Fc' and the feature matrix Fa horizontally (to expand the dimension);

[0101] For example: , ; Matrix after horizontal splicing = ;

[0102] The matrix after horizontal splicing is input into a neural network D, and the neural network D is used to extract the features of the matrix and output a feature matrix Fd (reconstructed image feature layer);

[0103] Specifically, the concatenated matrix [Fc' Fa] passes through the convolution layer, pooling layer, fully connected layer, and output layer of the neural network D in sequence to obtain a feature matrix Fd of size C×W×H.

[0104] Among them, in order to improve the accuracy, the number of convolutional layers of the neural network D is greater than 5;

[0105] The feature matrix Fd is then passed through the maximum pooling layer to obtain a weight matrix, which is consistent with the size of the feature matrix Fb (C×1×1). The weight matrix is ​​applied to the feature matrix Fd (reorganized image feature layer) to obtain the feature matrix Fd', that is, each channel is multiplied by its own weight.

[0106] The value of each element in the feature matrix Fd' is recorded as the gray value of the pixel point to generate a defect sample image;

[0107] ② Input the feature matrix Fd into the neural network E, which is used to extract the features of the feature matrix Fd, perform identification, and output a defect information matrix with the same dimension as the label (N rows × 6 columns);

[0108] Among them, in order to improve the accuracy, the number of convolutional layers of the neural network E is greater than 13;

[0109] The loss value is calculated using the similarity loss1 between the generated defect sample image and the normal image and the similarity loss2 between the defect information matrix and the label;

[0110] That is, loss=loss1+loss2;

[0111] 3) Determine whether the loss value has reached stability:

[0112] If not, switch to the next set of training samples and continue to step 2) to iterate the model parameters;

[0113] If yes, the deep learning model training is completed and the model parameters are saved.

[0114] Next, the method of using the trained deep learning model to generate defect sample images is as follows:

[0115] Set label I, which includes: the position coordinates, size information, type and severity of the defect to be generated; By setting label I, the defect position, type and severity can be customized.

[0116] In a set of defect images consistent with the defect type in label I, search for one or more defect images that are consistent with and / or similar to the severity level in the label;

[0117] Input any normal image, label I, and one or more defective images found (consistent with the number of defective images in the sample) into the trained deep learning model, and perform the processing of step ① to generate defective sample images;

[0118] The generated defect sample image complies with the information set in label I.

[0119] In order to increase the diversity of defect severity levels, it is preferably implemented that in label I, the defect severity level can be set within a range from the minimum level to the maximum level of the defect levels corresponding to the N types of defects.

[0120] That is to say, the severity level of the defect set in label I is not limited to the severity level range of a certain defect in the existing training samples. For example, in the training samples corresponding to scratches, there are three types of defect images with severity levels 1, 2, and 5;

[0121] Among the training samples corresponding to the pits, there are three defect images with severity levels 1, 2, and 3;

[0122] Among the training samples corresponding to the dirty points, there are four defect images with severity levels 2, 3, 4, and 5;

[0123] At this time, the minimum level to the maximum level of the defect levels corresponding to the N types of defects are 1 to 5; in label I, the severity levels of the three types of defects, scratches, pits, and dirt spots, can be set in the range of 1 to 5.

[0124] This facilitates increasing the diversity of generated defect sample images.

[0125] In specific implementation, in step 1), the severity level of the defect is evaluated as follows:

[0126] Perform the following processing on each defect image set to obtain the defect levels corresponding to N types of defects:

[0127] ① Calculate the grayscale mean G' of the defect area in a single defect image in the current defect image set;

[0128] Find the image area corresponding to the defect area in M ​​normal images and calculate the grayscale mean G of the found image area i , i=1,2...M; take M G i The median or mean is recorded as G'';

[0129] Calculate the difference between G'' and G';

[0130] ② Perform step ① on each defect image in the defect image set to obtain multiple difference values;

[0131] ③ Set the severity level of the defect according to the size of the difference.

[0132] More specifically, in step ③, the severity level of the defect is set according to the size of the difference, in the following two ways:

[0133] Method 1: Sort multiple differences from small to large, and divide each S difference into a level;

[0134] Method 2: Perform K nearest neighbor clustering on multiple differences to obtain K clustering results, and divide K severity levels according to the clustering results.

[0135] For example, K clustering results are obtained, and the cluster center of each clustering result is found. If the cluster center is a difference of 5, 0 to 10 are classified into a severity level; if the cluster center is 30, 25 to 30 are classified into a severity level.

[0136] This method trains the deep learning model to identify defects of different categories and different severity levels, so that the model can generate defect sample images with adjustable positions, variable types, and independently designed severity levels, which increases the diversity of defect sample images and has the characteristics of high accuracy and strong universality.

[0137] The foregoing description of the specific exemplary embodiments of the present invention is presented for the purpose of illustration and description. The foregoing description is not intended to be exhaustive, nor is it intended to limit the present invention to the precise form disclosed, and it is apparent that many changes and variations are possible in light of the above teachings. The exemplary embodiments are selected and described to explain the specific principles of the present invention and its practical application, so that other technicians in the field can realize and utilize the various exemplary embodiments of the present invention and its different selected forms and modified forms. The scope of the present invention is intended to be limited by the appended claims and their equivalents.

Claims

1. A defect sample image generation method, characterized in that: Generate defect sample images using a trained deep learning model, wherein the deep learning model includes neural network A, neural network B, neural network C, neural network D and neural network E; The dimensions of the output layers of neural network A, neural network C, and neural network D are the same; The number of channels in the output layer of neural network B and neural network C is the same; The dimensions of the output layers of neural network B and neural network E are the same; The training steps of the deep learning model are as follows: 1) Collect multiple images of the object to be tested in the same inspection scene, including various defect images and normal images; Defect images of different defect categories are stored separately to form N defect image sets; N represents the number of defect categories; Label each defect image, wherein the label includes the location coordinates, size information, type and severity level of the defect; For each label, in a set of defect images with the same type, find defect images with the same and / or similar severity levels as those in the label and mark them as associated images; A single normal image, a single label, and the associated images corresponding to the label are recorded as a group of training samples; different labels and different normal images are randomly combined to form multiple groups of training samples; 2) Input a set of training samples into the deep learning model and perform the following processing: ① The normal image is input into the neural network A, and the neural network A is used to extract the features of the normal image and output the feature matrix Fa corresponding to the normal image; The label is input into a neural network B, and the neural network B is used to extract label features and output a feature matrix Fb corresponding to the label; The defect image is input into a neural network C, and the neural network C is used to extract features of the defect image and output a feature matrix Fc corresponding to the defect image; Apply the feature matrix Fb to the feature matrix Fc to obtain the feature matrix Fc'; The feature matrix Fc' is horizontally spliced ​​with the feature matrix Fa and the matrix after the horizontal splicing is input into the neural network D, the neural network D is used to extract the features of the matrix and output the feature matrix Fd; The feature matrix Fd is then passed through the maximum pooling layer to obtain a weight matrix, and the weight matrix is ​​consistent with the feature matrix Fb in size; the weight matrix is ​​applied to the feature matrix Fd to obtain the feature matrix Fd'; The value of each element in the feature matrix Fd' is recorded as the gray value of the pixel point to generate a defect sample image; ② Input the feature matrix Fd into the neural network E, which is used to extract the features of the feature matrix Fd, perform identification, and output a defect information matrix with the same dimension as the label; The loss value is calculated using the similarity between the generated defect sample image and the normal image, and the similarity between the defect information matrix and the label; 3) Determine whether the loss value has reached stability: If not, switch to the next set of training samples and continue to step 2) to iterate the model parameters; If yes, the deep learning model training is completed and the model parameters are saved.

2. The defect sample image generation method according to claim 1, characterized in that: The method of generating defect sample images using the trained deep learning model is as follows: Setting label I, which includes: location coordinates, size information, type and severity of the defect to be generated; In a set of defect images consistent with the defect type in label I, search for one or more defect images that are consistent with and / or similar to the severity level in the label; Input any normal image, label I, and one or more defect images found into the trained deep learning model, perform the processing of step ①, and generate defect sample images; The generated defect sample image complies with the information set in label I.

3. The defect sample image generation method according to claim 2, characterized in that: In label I, the severity level of the defect can be set in the range of: the minimum level to the maximum level among the defect levels corresponding to the N types of defects.

4. The defect sample image generation method according to claim 1, characterized in that: In step 1), the severity of the defect is assessed as follows: Perform the following processing on each defect image set to obtain the defect levels corresponding to N types of defects: ① Calculate the grayscale mean G' of the defect area in a single defect image in the current defect image set; Find the image area corresponding to the defect area in M ​​normal images and calculate the grayscale mean G of the found image area i , i=1,2...M; take M G i The median or mean is recorded as G''; Calculate the difference between G'' and G'; ② Perform step ① on each defect image in the defect image set to obtain multiple difference values; ③ Set the severity level of the defect according to the size of the difference.

5. The defect sample image generation method according to claim 4, characterized in that: In step ③, the severity level of the defect is set according to the size of the difference. There are two ways: Method 1: Sort multiple differences from small to large, and divide each S difference into a level; Method 2: Perform K nearest neighbor clustering on multiple differences to obtain K clustering results, and divide K severity levels according to the clustering results.

6. The defect sample image generation method according to claim 1, characterized in that: When there are multiple defect images input into the neural network C, a weight value is set for each defect image in advance; Each defect image corresponds to a feature matrix, and the corresponding elements in multiple feature matrices are weighted and added to obtain the feature matrix Fc.

7. The defect sample image generation method according to claim 6, characterized in that: The weights are designed in the following way: The weight of defect images that are consistent with the severity level in the label is greater than the weight of defect images that are similar to the severity level in the label.

8. The defect sample image generation method according to claim 1, characterized in that: The network models selected by neural network A, neural network B, neural network C, neural network D and neural network E are not exactly the same.

9. The defect sample image generation method according to claim 1, characterized in that: Neural network A uses a convolutional neural network with more than 13 convolutional layers; Neural network B uses a convolutional neural network with more than 3 convolutional layers; Neural network C uses a convolutional neural network with more than 13 convolutional layers; Neural network D uses a convolutional neural network with more than 5 convolutional layers; The neural network E uses a convolutional neural network with more than 13 convolutional layers.

Images