A metallographic microstructure image generation method based on improved generative adversarial network

Abstract

The present invention provides a method for generating metallographic microstructure images based on an improved generative adversarial network, which belongs to the field of intelligent control of the microstructure of steel. The present invention uses an inverse residual network as a generator of a generative adversarial network to realize the conversion of the composition and process data of a one-dimensional metal material to a three-dimensional metallographic microstructure image. Compared with the traditional metallographic microstructure image acquisition process, the method of the present invention greatly reduces the metallographic microstructure image generation time. By using the generator in the improved generative adversarial network, small-size single-channel image data is converted into large-size three-channel metallographic microstructure image data, avoiding the time-consuming metallographic microstructure image acquisition process, and can help researchers observe the metallographic microstructure image corresponding to the given metal material composition and process data in advance, which is conducive to improving the production efficiency of steel enterprises.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent regulation of steel microstructure, and in particular relates to a metallographic microstructure image generation method based on an improved generative adversarial network. Background Art

[0002] Metallographic microstructure refers to the structure observed under an optical microscope or electron microscope on the ground surface or film of a metal sample treated by an appropriate method (such as corrosion). Metallographic microstructure image analysis is one of the main methods for studying the relationship between the properties, composition, organizational structure, and processing technology of metal materials, and plays a pivotal role in the field of materials science. The metallographic microstructure image of steel contains rich organizational morphology and distribution information, which has a decisive influence on the chemical and mechanical properties of steel materials.

[0003] At present, the main steps for obtaining the metallographic microstructure image of steel are as follows: (1) Sampling. When cutting steel, the entire wall thickness section of the inspection part should be included. For sections with larger wall thickness, several samples are allowed to be made. (2) Grinding. First, the cut sample is ground flat on a grinding wheel machine. When grinding, the force is light and even. Then, metallographic sandpaper of different particle sizes is used to gradually polish it. (3) Polishing. After the metallographic sample is polished, there are fine scratches and metal deformation disturbances on the surface, so polishing is performed. Polishing can remove the marks on the grinding surface and eliminate the deformation disturbance layer on the grinding surface. The polishing of the sample includes mechanical polishing, chemical polishing and electrolytic polishing. (4) Corrosion. In order to obtain the information of the microstructure of the polished metallographic sample, it must be properly corroded (such as using 4% nitric acid alcohol solution to corrode the metal sample) so that the microstructure can be correctly displayed. (5) Collection. The metallographic image of the prepared metallographic sample is collected through a metallographic microscope. The above-mentioned method of obtaining metallographic microstructure images is affected by the subjectivity of the inspectors. At the same time, this metallographic microstructure image acquisition process is not only time-consuming but also costly. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a method for generating metallographic microstructure images based on an improved generative adversarial network.

[0005] The technical solution of the present invention is:

[0006] 1. A method for generating metallographic microstructure images based on an improved generative adversarial network, characterized in that the method comprises the following steps:

[0007] Step 1: Read a metallographic microstructure image in the training set and define it as a three-dimensional matrix of size (C, W, H) according to the size of the image, expressed as X(C, W, H);

[0008] Step 2: Pass X through the discriminator to obtain the feature map M 1 ~M 5 , and the feature map M 5 After the Sigmoid function is applied, the probability of the metallographic microstructure image being true is outputted;

[0009] Step 3: Use a binary cross entropy loss function to calculate the loss value between the probability that the metallographic microstructure image obtained in step 2 is true and the true probability 1, and perform a reverse transfer, so as to iteratively optimize the model parameters in the discriminator to obtain a discriminator model that is iteratively trained once;

[0010] Step 4: In the generator, a data conversion module is applied to convert the composition and process data of the one-dimensional metal material corresponding to the metallographic microstructure image in the training set into a single-channel three-dimensional matrix G, expressed as: (1, W′, H′);

[0011] Step 5: Add the residual unit to the generator to obtain an improved generator, and pass the matrix G through the improved generator to obtain the feature map M 6 ~M 12 ;

[0012] Step 6: The feature map M obtained in step 5 12 Input to the discriminator obtained in step 3 after one iteration of training, and 12 Execute steps 2 and 3 to get the loss value and perform reverse propagation;

[0013] Step 7: Execute steps 1 to 6 within the set number of iterations to obtain the trained generator;

[0014] Step 8: Input the composition and process data of the metal materials in the test set into the trained generator to generate metallographic microstructure images.

[0015] Further, according to the metallographic microstructure image generation method based on the improved generative adversarial network, the X is passed through the discriminator to obtain a feature map M 1 ~M 5 The method is: change the number of feature map channels by convolution operation on X to obtain M 1 (C 1 ,W,H); then the feature map size is changed by the size and step size of the convolution kernel in the convolution layer to obtain the feature map M 2 ~M 5 ; Feature map M 2 ~M 5 The sizes are: M 5 (1,1,1).

[0016] Further, according to the metallographic microstructure image generation method based on the improved generative adversarial network, the X is passed through the discriminator to obtain a feature map M 1 ~M 5 The method comprises the following steps:

[0017] Step 2.1: Extract features from X, change the number of channels of X through convolution operation, and then obtain feature map M through data standardization and activation operation. 1 (C 1 ,W,H);

[0018] Step 2.2: M 1 Perform feature extraction and change M through convolution operation 1 The size of the feature map, and then the feature map is obtained by data normalization and activation operation

[0019] Step 2.3: M 2 Perform feature extraction and change M through convolution operation 2 The size of the feature map, and then the feature map is obtained by data normalization and activation operation

[0020] Step 2.4: M 3 Perform feature extraction and change M through convolution operation 3 The size of the feature map, and then the feature map is obtained by data normalization and activation operation

[0021] Step 2.5: M 4 Perform feature extraction and change M through convolution operation 4 The size of the feature map is then normalized and activated to obtain the feature map M. 5 (1,1,1);

[0022] Step 2.6: Apply the classification function to the feature map M 5 Calculation is performed, and the classification function outputs the probability that the metallographic microstructure image is true.

[0023] Furthermore, according to the metallographic microstructure image generation method based on the improved generative adversarial network, the classification function is a Sigmoid function.

[0024] Further, according to the metallographic microstructure image generation method based on the improved generative adversarial network, the step 5 includes the following steps:

[0025] Step 5.1: Extract features from G, and change the number of feature map channels through the convolution operation of G to obtain the feature map M 6 (C 2 ,W′,H′);

[0026] Step 5.2: M 6 Perform feature extraction and convert M 6 After upsampling, its size is changed to get Then After two serial residual units, the change The number of feature map channels is obtained

[0027] Step 5.3: M 7 Perform feature extraction and convert M 7 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0028] Step 5.4: M 8 Perform feature extraction and convert M 8 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0029] Step 5.5: M 9 Perform feature extraction and convert M 9 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0030] Step 5.6: M 10 Perform feature extraction and convert M 10 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0031] Step 5.7: M 11 Perform feature extraction and apply convolution operations to change the size and number of channels of the feature map to meet the discriminator's input data format requirements, and finally obtain M 12 (C1 ,W,H).

[0032] Furthermore, according to the metallographic microstructure image generation method based on the improved generative adversarial network, each of the residual units is composed of a convolutional layer, data normalization and an activation function.

[0033] Furthermore, according to the metallographic microstructure image generation method based on the improved generative adversarial network, the activation function in the residual unit is a ReLU function.

[0034] Compared with the prior art, the advantages and beneficial effects of this invention are:

[0035] (1) The inverse residual network is used as the generator of the generative adversarial network to avoid the problem of gradient vanishing or gradient exploding during training due to the increase in the number of network model layers. At the same time, compared with the traditional residual network, the inverse residual network realizes the transformation of the feature map size from small to large.

[0036] (2) The improved generative adversarial network is used to realize the conversion from one-dimensional metal material composition and process data to three-dimensional metallographic microstructure images. This is conducive to improving the generative adversarial network to learn the global information between the composition and process data of metal materials, thereby improving the accuracy of metallographic microstructure image generation.

[0037] (3) Compared with the traditional metallographic microstructure image acquisition process, the method of the present invention greatly reduces the generation time of metallographic microstructure images. By using the generator in the improved generative adversarial network, small-size (such as 1×32×32) single-channel image data is converted into large-size (such as 3×1024×1024) three-channel metallographic microstructure image data, avoiding the time-consuming metallographic microstructure image acquisition process.

[0038] (4) It can help researchers observe the metallographic microstructure images corresponding to the given metal material composition and process data in advance, which is conducive to improving the production efficiency of steel enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the process of the metallographic microstructure image generation method based on the improved generative adversarial network in this embodiment;

[0040] Figure 2 A schematic diagram of a discriminator network in an improved generative adversarial network according to this embodiment;

[0041] Figure 3 A schematic diagram of a generator network in an improved generative adversarial network according to this embodiment;

[0042] Figure 4 Schematic diagram of the structure of a single residual unit in this implementation mode;

[0043] Figure 5 This is an example of a metallographic microstructure image generated in this embodiment. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0045] Figure 1 is a flow chart of a method for generating metallographic microstructure images based on an improved generative adversarial network according to the present invention, such as Figure 1 As shown, the metallographic microstructure image generation method based on the improved generative adversarial network includes the following steps:

[0046] Step 1: Read a metallographic microstructure image in the training set and define it as a three-dimensional matrix of size (C, W, H) according to the size of the image, expressed as X(C, W, H).

[0047] In the embodiment, one metallographic microstructure image in the training set is read, and the read metallographic microstructure image is represented by a three-dimensional matrix X(3,1024,1024), wherein the number 3 in the first dimension represents that the number of channels of the metallographic microstructure image is 3; and the numbers 1024 in the second and third dimensions represent the size of the metallographic microstructure image.

[0048] Step 2: Pass the metallographic microstructure image X through the discriminator to obtain the feature map M 1 ~M 5 , and the obtained M 5 After the S-shaped growth curve function (Sigmoid function), the probability of the metallographic microstructure image being true is output;

[0049] In this embodiment, the read metallographic microstructure image X is passed through the discriminator to obtain a feature map M 1 ~M 5 The method is: change the number of feature map channels by convolution operation on X to obtain M 1 (C 1 ,W,H); then the feature map size is changed by the size and step size of the convolution kernel in the convolution layer to obtain the feature map M 2 ~M 5 . M 2 ~M 5 The sizes of the feature maps are: M 5 (1,1,1). Finally, the obtained M 5The probability that the metallographic microstructure image is true is obtained after the S-shaped growth curve function, namely the Sigmoid function.

[0050] Figure 2 This is a schematic diagram of the network model of the discriminator, where each part consists of a convolution operation, a data normalization operation, and an activation function;

[0051] In the embodiment, the metallographic microstructure image X is passed through the discriminator to obtain a feature map M 1 ~M 5 The specific process of calculating and classifying probability includes the following steps 2.1 to 2.6.

[0052] Step 2.1: Extract features from X, change the number of channels of X through convolution operation, and then obtain feature map M through data standardization and activation operation. 1 (C 1 ,W,H).

[0053] In the embodiment, the number of convolution kernels of the convolution operation described in this step is 32, the size is 3×3, the step length is 1, and the padding is 1. First, the convolution operation is applied to extract features of X, and then the feature map after the convolution operation is subjected to data standardization and activation operations to obtain the feature map M 1 (32,1024,1024).

[0054] Step 2.2: M 1 Perform feature extraction and change M through convolution operation 1 The size of the feature map, and then the feature map is obtained by data normalization and activation operation

[0055] In the embodiment, the convolution operation described in this step has 64 convolution kernels, a size of 3×3, a step size of 2, and a padding of 1. First, the convolution operation is applied to M 1 Perform feature extraction, and then perform data standardization and activation operations on the feature map obtained after the convolution operation to obtain the feature map M 2 (64,512,512).

[0056] Step 2.3: M 2 Perform feature extraction and change M through convolution operation 2 The size of the feature map, and then the feature map is obtained by data normalization and activation operation

[0057] In the embodiment, the number of convolution kernels in the convolution operation described in this step is 128, the size is 3×3, the step length is 2, and the padding is 1. First, the convolution operation is applied to M 2 Perform feature extraction, and then perform data standardization and activation operations on the feature map after the convolution operation to obtain the feature map M3 (128,256,256).

[0058] Step 2.4: M 3 Perform feature extraction and change M through convolution operation 3 The size of the feature map, and then the feature map is obtained by data normalization and activation operation

[0059] In the embodiment, the convolution operation described in this step has 256 convolution kernels, a size of 3×3, a step size of 2, and a padding of 1. First, the convolution operation is applied to M 3 Perform feature extraction, and then perform data standardization and activation operations on the feature map after the convolution operation to obtain the feature map M 4 (256,128,128).

[0060] Step 2.5: M 4 Perform feature extraction and change M through convolution operation 4 The size of the feature map is then normalized and activated to obtain the feature map M. 5 (1,1,1);

[0061] In the embodiment, the convolution operation described in this step has 1 convolution kernel, 128×128 size, 1 step length, and 0 padding. First, the convolution operation is applied to M 4 Perform feature extraction, and then perform data standardization and activation operations on the feature map after the convolution operation to obtain the feature map M 5 (1,1,1).

[0062] Step 2.6: Apply the classification function to the feature map M 5 Calculation is performed, and the classification function outputs the probability that the metallographic microstructure image is true;

[0063] In an embodiment, a classification function Sigmoid (S-shaped growth curve) function is applied to the feature map M 5 The calculation is performed to obtain the probability that the metallographic microstructure image is true. The S-shaped growth curve function is shown below.

[0064]

[0065] Step 3: Use the binary cross entropy loss function to calculate the loss value between the probability obtained in step 2 and the true probability 1, and perform a reverse transfer to iteratively optimize the model parameters in the discriminator to obtain a discriminator model that is iteratively trained once;

[0066] In an embodiment, a binary cross entropy loss function is applied to calculate M 5The loss value between the true probability with the same shape and value of 1 is back-propagated in the discriminator to iteratively optimize the model parameters in the discriminator. The binary cross entropy loss function is shown below.

[0067]

[0068] Where n represents the number of samples, y i represents the label of the i-th sample, z i It represents the probability that the i-th sample is predicted as a positive example.

[0069] Step 4: In the generator, a data conversion module is applied to convert the composition and process data of the one-dimensional metal material corresponding to the metallographic microstructure image in the training set into a single-channel three-dimensional matrix G, expressed as: (1, W′, H′);

[0070] In an embodiment, 32 parameter data in the composition and process of the metal material are selected and input into the data conversion module in the generator. The 32 parameter data include: carbon, silicon, manganese, phosphorus, sulfur, nitrogen, niobium, vanadium, titanium, aluminum, copper, chromium, nickel, cobalt, molybdenum, boron, soaking temperature, rough rolling 1 pass temperature, rough rolling 2 pass temperature, rough rolling 3 pass temperature, rough rolling 4 pass temperature, rough rolling 5 pass temperature, rough rolling 6 pass temperature, rough rolling 7 pass temperature, finishing 1 pass temperature, finishing 2 pass temperature, finishing 3 pass temperature, finishing 4 pass temperature, finishing 5 pass temperature, finishing 6 pass temperature, finishing 7 pass temperature, coiling temperature. In the data conversion module, first establish a two-dimensional matrix G of size 32x32 and all values ​​are 1, and the 32 components and process data of the metal material are repeatedly filled into the matrix G in sequence after data normalization, and then the matrix G is dimensionally increased to the matrix G (1, 32, 32). The first dimension is 1, representing the number of channels of matrix G, and the second and third dimensions are both 32, indicating the size of matrix G.

[0071] Step 5: Add the residual unit to the generator to obtain an improved generator, and pass the matrix G through the improved generator to obtain the feature map M 6 ~M 12 ;

[0072] In the embodiment, the matrix G is passed through the generator to obtain the feature map M 6 ~M 12 The method is as follows: first, G is convolved to change the number of feature map channels, and the feature map M is obtained. 6 (C 2 ,W′,H′); then M i After upsampling, M is changed i The size of the feature map is obtained Then After two residual units, the feature map M is obtained i+1 , where i is the index of the feature map, i = 6, 7, 8, 9, 10, 11, the M i+1 They are M 12 (C 1 , W, H).

[0073] Figure 3 is a schematic diagram of the structure of the generator, where except for the first two modules and the last module, each of the remaining modules consists of upsampling and two residual units, where each residual unit consists of a convolutional layer, data normalization and an activation function. Figure 4 FIG. 1 is a schematic diagram showing the structure of a residual unit.

[0074] Step 5.1: Extract features from G, and change the number of feature map channels through the convolution operation to obtain the feature map M 6 (C 2 ,W′,H′);

[0075] In the embodiment, the number of convolution kernels in the convolution operation described in this step is 1024, the size is 3×3, the step length is 1, and the padding is 0. Convolution operation, data normalization operation and activation operation are performed on G to obtain a feature map M 6 (1024,32,32).

[0076] Step 5.2: M 6 Perform feature extraction and convert M 6 After upsampling, its size is changed to get Then After two serial residual units, the change The number of feature map channels is obtained

[0077] In the embodiment: the kernel size of all upsampling operations is 2×2, and the step size is 2; the convolution kernel size in all residual units is 3×3, the step size is 1, and the padding is 1. The difference is that the number of convolution kernels in the convolution layer is different, that is, the number of channels of the output feature map of each residual unit; the activation function in all residual units is the ReLU (rectified linear unit) function, and its expression is: f(x) = max(0, x). In the embodiment, M 6 After upsampling, the feature map size is changed, and we get Then After two residual units with 512 convolution kernels, the feature map M is obtained. 7 (512,64,64).

[0078] Step 5.3: M7 Perform feature extraction and convert M 7 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0079] In the embodiment, M 7 After upsampling, the size of the feature map is changed, and we get Then After two residual units with 256 convolution kernels, the feature map M is obtained. 8 (256,128,128).

[0080] Step 5.4: M 8 Perform feature extraction and convert M 8 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0081] In the embodiment, M 8 After upsampling, the size of the feature map is changed, and we get Then After two residual units with 128 convolution kernels, the feature map M is obtained. 9 (128,256,256).

[0082] Step 5.5: M 9 Perform feature extraction and convert M 9 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0083] In the embodiment, M 9 After upsampling, the size of the feature map is changed, and we get Then After two residual units with 64 convolution kernels, the feature map M is obtained. 10 (64,512,512).

[0084] Step 5.6: M 10Perform feature extraction and convert M 10 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained

[0085] In the embodiment, M 10 After upsampling, the size of the feature map is changed, and we get Then After two residual units with 32 convolution kernels, the feature map M is obtained. 11 (32,1024,1024).

[0086] Step 5.7: M 11 Perform feature extraction and apply convolution operations to change the size and number of channels of the feature map to meet the discriminator's requirements for the format of input data, and finally obtain M 12 (C 1 ,W,H);

[0087] In an embodiment, the convolution operation described in this step has 3 convolution kernels, a size of 3×3, a step size of 1, and a padding of 1. 11 Perform feature extraction and apply convolution operations to change the number and size of feature map channels to meet the discriminator's input data format requirements, and finally obtain M 12 (3,1024,1024).

[0088] Step 6: The feature map M obtained in step 5 12 Input to the discriminator obtained in step 3 after one iteration of training, and 12 Execute steps 2 and 3 to get the loss value and perform reverse propagation.

[0089] In the embodiment, the discriminator M trained once in step 3 is used. 12 (3,1024,1024) Execute steps 2 and 3 to obtain the loss value, and then iteratively optimize the generator parameters to obtain a generator that is iteratively trained once.

[0090] Step 7: Execute steps 1 to 6 within the set number of iterations epoch to obtain the trained generator.

[0091] In the embodiment, the number of iterations is set to 10,000, and within this number of iterations, steps 1 to 6 are continuously and repeatedly executed to finally obtain a trained generator model.

[0092] Step 8: Use the generator trained in step 7 to read the metal material composition and process data in the test set, realize the conversion from data to metallographic microstructure image, and thus generate a metallographic microstructure image.

[0093] In the embodiment, the composition and process data of the metal materials in the test set are input into the trained generator to generate the metallographic microstructure image. Figure 5 shown.

[0094] Obviously, the above embodiments are only some embodiments of the present invention, rather than all embodiments. The above embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention. Based on the above embodiments, all other embodiments obtained by those skilled in the art without making creative work, that is, all modifications, equivalent substitutions and improvements made within the spirit and principles of this application, fall within the protection scope required by the present invention.

Claims

1. A metallographic microstructure image generation method based on improved generative adversarial network, It is characterized in that The method comprises the following steps: Step 1: Read a metallographic microstructure image in the training set and define it as a three-dimensional matrix of size (C, W, H) according to the size of the image, expressed as X(C, W, H); Step 2: Pass X through the discriminator to obtain the feature map M 1 ~M 5 , and pass the feature map M 5 through the Sigmoid function and output the probability that the metallographic microstructure image is true; Step 3: Use a binary cross entropy loss function to calculate the loss value between the probability that the metallographic microstructure image obtained in step 2 is true and the true probability 1, and perform a reverse transfer, so as to iteratively optimize the model parameters in the discriminator to obtain a discriminator model that is iteratively trained once; Step 4: In the generator, a data conversion module is applied to convert the composition and process data of the one-dimensional metal material corresponding to the metallographic microstructure image in the training set into a single-channel three-dimensional matrix G, expressed as: (1, W′, H′); Step 5: Add the residual unit to the generator to obtain an improved generator, and pass the matrix G through the improved generator to obtain the feature map M 6 ~M 12 ; Step 6: The feature map M obtained in step 5 12 Input to the discriminator obtained in step 3 after one iteration of training, and 12 Execute steps 2 and 3 to get the loss value and perform reverse propagation; Step 7: Execute steps 1 to 6 within the set number of iterations to obtain the trained generator; Step 8: Input the composition and process data of the metal materials in the test set into the trained generator to generate metallographic microstructure images.

2. The metallographic microstructure image generation method based on the improved generative adversarial network according to claim 1, It is characterized in that The X passes through the discriminator to obtain the feature map M 1 ~M 5 The method is: change the number of feature map channels by convolution operation on X to obtain M 1 (C 1 , W, H); then the feature map size is changed by the size and step size of the convolution kernel in the convolution layer to obtain the feature map M 2 ~M 5 ; Feature map M 2 ~M 5 The sizes are: M 5 (1, 1, 1).

3. The metallographic microstructure image generation method based on the improved generative adversarial network according to claim 2, It is characterized in that The above-mentioned X passes through a discriminator to obtain a feature map M 1 ~M 5 The method includes the following steps: Step 2.1: Extract features from X, change the number of channels of X through convolution operation, and then obtain feature map M through data standardization and activation operation. 1 (C 1 , W, H); Step 2.2: For M 1 Perform feature extraction and change M through convolution operation 1 The size of the feature map, and then the feature map is obtained by data normalization and activation operation Step 2.3: M 2 Perform feature extraction and change M through convolution operation 2 The size of the feature map, and then the feature map is obtained by data normalization and activation operation Step 2.4: M 3 Perform feature extraction and change M through convolution operation 3 The size of the feature map, and then the feature map is obtained by data normalization and activation operation Step 2.5: M 4 Perform feature extraction and change M through convolution operation 4 The size of the feature map is then normalized and activated to obtain the feature map M. 5 (1, 1, 1); Step 2.6: Apply the classification function to the feature map M 5 Calculation is performed, and the classification function outputs the probability that the metallographic microstructure image is true.

4. The metallographic microstructure image generation method based on the improved generative adversarial network according to claim 3, It is characterized in that The classification function is a Sigmoid function.

5. The metallographic microstructure image generation method based on the improved generative adversarial network according to claim 3, It is characterized in that The step 5 comprises the following steps: Step 5.1: Extract features from G, and change the number of feature map channels through the convolution operation to obtain the feature map M 6 (C 2 , W′, H′); Step 5.2: M 6 Perform feature extraction and convert M 6 After upsampling, its size is changed to get Then After two serial residual units, the change The number of feature map channels is obtained Step 5.3: M 7 Perform feature extraction and convert M 7 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained Step 5.4: M 8 Perform feature extraction and convert M 8 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained Step 5.5: M 9 Perform feature extraction and convert M 9 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained Step 5.6: M 10 Perform feature extraction and convert M 10 After upsampling, its size is changed to get Then After two serial residual units for feature extraction, the change The number of feature map channels is obtained Step 5.7: M 11 Perform feature extraction and apply convolution operations to change the size and number of channels of the feature map to meet the discriminator's requirements for the format of input data, and finally obtain M 12 (C 1 , W, H).

6. The method for generating metallographic microstructure images based on an improved generative adversarial network according to claim 5, It is characterized in that Each of the residual units consists of a convolutional layer, data normalization and an activation function.

7. The method for generating metallographic microstructure images based on an improved generative adversarial network according to claim 5, It is characterized in that The activation function in the residual unit is the ReLU function.

Images