A metallographic structure recognition method based on multi-objective fractal evolution network

Abstract

This invention discloses a metallographic structure recognition method based on a multi-objective fractal evolutionary network, relating to the field of steel material image recognition. The invention utilizes a data transfer method based on fractal dimension deviation. First, it transfers data from fabric texture images with fractal characteristics, expanding the metallographic structure image data volume. Second, based on the metallographic structure data, it generates new data using a generative adversarial network (GAN) to further expand the metallographic structure image data. The expanded data is then input into a fractal network for training to achieve metallographic structure recognition. During training, a multi-objective differential evolutionary algorithm is used to optimize the fractal network, obtaining a Pareto optimal solution. Finally, a distance-based method is used to determine Knee points to obtain a metallographic structure recognition model, thereby enabling metallographic structure recognition of input metallographic structure images in practical applications. This method solves the problems of difficult metallographic structure recognition, limited data volume, and unbalanced data types, improving the accuracy of metallographic structure recognition.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology for steel materials, and specifically to a metallographic structure recognition method based on a multi-objective fractal evolution network. Background Technology

[0002] Metallographic structure reflects the specific internal morphology of metallic and alloy materials and is closely related to their comprehensive properties (mechanical properties, thermal properties, etc.). Metallographic analysis is an important tool for studying the internal structure and performance testing of metallic and alloy materials, playing a crucial role in the field of materials science. Metallographic structure identification is the prerequisite and foundation of metallographic analysis, and is the core and key to automated quantitative analysis of the internal structure of metallographic structures.

[0003] Currently, in metallographic examination, the metallographic structure of materials is identified and evaluated by inspectors through visual observation using a metallographic electron microscope. This requires inspectors to have extensive professional knowledge and field experience, and the accuracy of the results is greatly affected by human factors. Furthermore, the metallographic preparation process is difficult, requiring the acquisition of a large number of samples, making it an extremely time-consuming and labor-intensive project, posing significant challenges to production and research. In recent years, with the continuous development of computer image recognition technology, new technical avenues have been opened up for research in the field of materials science. Many researchers utilize computer-aided metallographic analysis, employing image processing techniques to identify and quantify the characteristics of metallographic structures. However, metallographic structure identification based on traditional image processing techniques requires feature extraction combined with expert experience. Human intervention can lead to the loss of some important information and low efficiency, thereby reducing the accuracy of identification.

[0004] To address the aforementioned issues, some researchers have proposed deep learning-based methods for metallographic structure identification, which have to some extent reduced the error rate of manually classifying metallographic structure images and improved the accuracy of metallographic structure identification. However, deep learning-based metallographic structure identification requires a large amount of data, while the metallographic preparation process of materials is complex, resulting in a small number of data samples. Furthermore, the types, quantities, and distribution of metallographic structures are strongly correlated with the material type and application scenario, leading to an imbalance in the metallographic image data samples. Consequently, this method cannot achieve a high accuracy rate in metallographic structure identification. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention provides a metallographic structure identification method based on a multi-objective fractal evolution network, which aims to solve the problems of difficult metallographic structure identification, small amount of data, and unbalanced data types, thereby improving the accuracy of deep learning-based metallographic structure identification methods.

[0006] The technical solution of this invention is:

[0007] A metallographic structure identification method based on a multi-objective fractal evolution network, comprising the following steps:

[0008] Collect images of metallographic structures of different categories, establish an original metallographic structure dataset, and divide the dataset into a training dataset and a test dataset according to a preset ratio;

[0009] Find the fractal dimension D of each metallographic image in the training dataset. B ;

[0010] Create a fabric texture image dataset and calculate the fractal dimension of each image in the dataset.

[0011] Based on D B and Transfer fabric texture images to the training dataset to augment the sample data size of the original training dataset.

[0012] The data from the expanded training dataset is input into the fractal network, and a multi-objective differential evolution algorithm is used to optimize the fractal network during the training process, gradually finding the Pareto optimal solution.

[0013] The Knee point is determined from the Pareto optimal solution to obtain the metallographic structure identification model;

[0014] The metallographic structure recognition model was trained using the expanded training dataset to obtain a well-trained metallographic structure recognition model.

[0015] The trained metallographic structure recognition model is used to perform metallographic structure recognition on the input metallographic structure image.

[0016] Furthermore, the fractal dimension D of each metallographic image was calculated using the box counting method. B The fractal dimension of each fabric texture image was calculated using box counting.

[0017] Furthermore, the fractal dimension D is calculated. B and The absolute value Δ of the difference is used to transfer the fabric texture images corresponding to Δ that are less than a preset threshold to the training dataset, thereby expanding the sample data volume of the training dataset.

[0018] Furthermore, the method also includes the steps of: generating new data that is highly similar to metallographic data using a generative adversarial network based on the original training dataset, adding the new data to the training dataset after the first sample data augmentation, and performing a second sample data augmentation on the training dataset.

[0019] Furthermore, with minimizing the time complexity and recognition error rate of the fractal network as optimization objectives, a Pareto optimal solution that balances these two objectives is gradually found during the optimization of the fractal network using a multi-objective differential evolution algorithm.

[0020] Furthermore, minimizing the time complexity of the fractal network can be viewed as minimizing the time complexity of the convolutional layer.

[0021] Furthermore, a distance-based method is used to determine the Knee point from the Pareto optimal solution.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) This invention utilizes a deep neural network—a fractal network—with sub-paths of different lengths for feature extraction to identify the metallographic structure of materials, and the network structures have a certain degree of similarity. Furthermore, it uses fractal dimension bias to select data from texture image data that are similar to the metallographic structure image of the material to expand the metallographic structure image of the material. Additionally, it employs a generative adversarial network to train the metallographic structure data to generate new data, thus expanding the metallographic structure dataset of the material and solving the problems of insufficient metallographic structure data, unbalanced data distribution, and inaccurate identification.

[0024] (2) The fractal network was optimized by multi-objective differential evolution algorithm, which balanced the two optimization objectives of model complexity and recognition accuracy. The Knee point selection strategy was used to determine the high-performance metallographic structure recognition model from the Pareto optimal solution, which enhanced the adaptability and generalization ability of the model.

[0025] (3) By verifying the collected metallographic structure datasets of different materials, the method of the present invention can significantly improve the accuracy and robustness of the metallographic structure identification results. Attached Figure Description

[0026] Figure 1 This is a flowchart of the metallographic structure identification method using a multi-objective fractal evolution network in this embodiment;

[0027] Figure 2 Images showing the metallographic structures of different steel materials used in this embodiment, where (a) is austenite; (b) is pearlite; (c) is martensite; (d) is cementite; and (e) is ferrite.

[0028] Figure 3 This is a flowchart of the fractal dimension calculation in this embodiment;

[0029] Figure 4These are texture images of different materials used in this embodiment, wherein (a) is a texture image of blanket-1; (b) is a texture image of blanket-2; (c) is a texture image of canvas-1; (d) is a texture image of ceiling-1; (e) is a texture image of ceiling-2; (f) is a texture image of mat-1; (g) is a texture image of floor-1; (h) is a texture image of floor-2; (i) is a texture image of lawn-1; and (j) is a texture image of lentil-1.

[0030] Figure 5 This is a diagram of the fractal network structure in this embodiment;

[0031] Figure 6 This is a schematic diagram of the fractal block expansion rules in this embodiment;

[0032] Figure 7 This is a flowchart of the multi-objective differential evolution algorithm in this embodiment;

[0033] Figure 8 This is a schematic diagram illustrating the distance-based method for determining the Knee point in this embodiment. Detailed Implementation

[0034] To facilitate understanding of this application, a more comprehensive description of this application will be provided below with reference to the accompanying drawings.

[0035] The core idea of ​​this invention is as follows: This invention proposes a metallographic structure recognition method based on a multi-objective fractal evolutionary network to improve the accuracy of existing deep learning-based metallographic structure recognition methods and solve the problems of insufficient metallographic structure data and unbalanced data category distribution. Since metallographic structures possess fractal characteristics, this invention proposes a data transfer method based on fractal dimension deviation to transfer data from fabric texture images with fractal characteristics, thus expanding the amount of metallographic structure image data for the first time. Next, based on the metallographic structure data, new data is generated through a generative adversarial network to further expand the metallographic structure image data. Then, the expanded data is input into the fractal network for training to achieve metallographic structure recognition. During training, a multi-objective differential evolutionary algorithm is used to optimize the fractal network, obtaining Pareto optimal solutions, each corresponding to a fractal network with a different structure. Finally, a distance-based method is used to determine Knee points to obtain a metallographic structure recognition model, thereby enabling metallographic structure recognition of input metallographic structure images in practical applications. Figure 1 This is a flowchart of the metallographic structure identification method based on a multi-objective fractal evolution network in this embodiment. Figure 1 As shown, the metallographic structure identification method based on multi-objective fractal evolution network includes the following steps:

[0036] Step 1: Collect metallographic images of different categories to establish an original metallographic dataset; divide the original metallographic dataset into a training dataset and a test dataset according to a preset ratio; in this embodiment, the collected metallographic images are labeled with metallographic categories, and some image sample data are shown below. Figure 2 As shown, the metallographic image sample data collected in this embodiment includes 56 austenite, 27 pearlite, 56 martensite, 12 cementite, and 15 ferrite, totaling 166 samples constituting the original metallographic dataset. In this embodiment, these 166 samples are randomly divided into the training dataset and the test dataset at a ratio of 4:1, but it is necessary to ensure that both the training dataset and the test dataset contain 5 types.

[0037] Step 2: Calculate the fractal dimension D of each metallographic image in the training dataset using box counting. B ;

[0038] In this embodiment, the box counting method is used to calculate the fractal dimension D of each metallographic image in the training dataset. B Methods such as Figure 3 As shown, it includes the following steps:

[0039] Step 2.1: Convert the input metallographic image X∈ H×W×3 Set the resolution to 450*450 and convert it to an 8-bit grayscale image;

[0040] Step 2.2: Convolve the Laplacian operator {[0,-1,0],[-1,5,-1],[0,-1,0]} with the grayscale image obtained in Step 2.1 to achieve the purpose of sharpening details;

[0041] Step 2.3: Apply the feature matrix X obtained in Step 2.2 C ∈ H×W×3 The standardization process is performed, and then the standardized feature matrix is ​​binarized to obtain matrix data containing only 0 and 1.

[0042] The standardization process is as follows: Calculate the mean μ of the training dataset after processing in step 2.2. 3 With variance σ∈ 3 Formula (1) is used for X C ∈ H×W×3 Standardization processing is performed, that is:

[0043]

[0044] Where H represents the height of the input image, W represents the width of the input image, and Y represents the image after normalization.

[0045] Step 2.4: Divide the matrix data obtained in Step 2.3 into 90*90 matrix data to improve computational efficiency;

[0046] Step 2.5: Set an appropriate box length ε (ε < 90) to segment the matrix data after cutting in Step 1.4;

[0047] Step 2.6: Calculate the number of boxes N(ε) contained in the matrix data according to different box lengths ε, so as to obtain the data set (ε, N(ε));

[0048] Step 2.7: Perform linear fitting on the data set (ε, N(ε)) obtained in Step 2.6, and obtain the slope D. B This is the fractal dimension of the metallographic image, calculated using the following formula:

[0049]

[0050] Step 3: Establish a fabric texture image dataset, and use box counting to calculate the fractal dimension of each image in the fabric texture image dataset.

[0051] In this embodiment, each fabric texture image in the fabric texture image dataset is categorized and labeled, including ten categories: ceiling-1, ceiling-2, blanket-1, blanket-2, floor-1, floor-2, canvas-1, mat-1, lawn-1, and lentil-1. Each category contains 160 texture images, resulting in a total of 1600 texture images in the fabric texture image dataset. Some texture image data is shown below. Figure 4 As shown; then, the fractal dimension of each image is calculated using box counting.

[0052] Step 4: Based on D B and The fabric texture images are transferred to the training dataset established in step 1 to perform the first sample data augmentation on the training dataset; the specific method is as follows:

[0053] The fractal dimension D is calculated according to equation (3). B and The absolute value Δ of the difference is used to transfer the fabric texture images corresponding to Δ that are less than a preset threshold to the training dataset, thereby expanding the sample data volume of the training dataset.

[0054]

[0055] In this embodiment, the threshold is set to 0.1 based on experience and experimental results. If Δ is less than the threshold of 0.1, it is determined that the corresponding metallographic image and fabric texture image are similar and meet the data migration conditions. Then, the fabric texture image that meets the conditions is migrated to the training dataset.

[0056] Step 5: Based on the original training dataset, use a Generative Adversarial Network (GAN) to generate new data that is highly similar to the metallographic data, and add the new data to the training dataset after the first sample data augmentation. Perform a second sample data augmentation on the training dataset to obtain the augmented dataset.

[0057] In this embodiment, the sample data in the original training dataset is first standardized. Then, the standardized data is input into the generative adversarial network for training. Through the interaction between the generator and the discriminator, the network is continuously judged and optimized to generate new data that is highly similar to the metallographic data. Finally, the new data generated by the generative adversarial network is stored, and the new data is added to the training dataset after the first sample data expansion. The training dataset is then expanded a second time to solve the problems of insufficient metallographic data and unbalanced data distribution, thus obtaining an expanded dataset.

[0058] Step 6: Input the data from the expanded dataset into the fractal network, and use a multi-objective differential evolution algorithm to optimize the two objectives of the fractal network's recognition accuracy and complexity during the training process, gradually finding the Pareto optimal solution that balances these two objectives;

[0059] Fractal network structures such as Figure 5 As shown, this network is an extremely deep neural network independent of residuals, composed of sub-path blocks of varying lengths, and includes pooling layers, fully connected layers, and softmax layers. To improve network performance and avoid overfitting, a path dropping method is used to regularize the fractal network, modifying the interactions between network layers. The construction of the fractal network includes the following steps:

[0060] (1) Construct the base layer of the fractal network. In this embodiment, a single convolutional unit is used as the base layer, which can be represented as F. C (z), where c represents the number of columns, z represents the input, and c = 1 represents a base layer;

[0061] First, convolutional units are constructed. A convolutional unit consists of a convolutional layer, a batch normalization layer, and an activation function. The convolutional layer is used for feature extraction, and the batch normalization layer normalizes the data after convolution, reducing the learning difficulty for the network. Finally, ReLU is chosen as the activation function.

[0062] Convolutional layers are used for feature extraction, and their mathematical expression is:

[0063]

[0064] in, These represent the input and output of the current convolutional layer, respectively. J represents the number of convolutional kernels in the current convolutional layer. j It is the number of channels of the j-th convolutional kernel. Let represent the weight of channel i of the j-th convolutional kernel in layer l, and This represents the bias of the j-th convolutional kernel in the l-th layer.

[0065] Next, a batch normalization layer is used to normalize the output data processed by the current convolutional layer, reducing the learning difficulty of the network.

[0066]

[0067] Where ò represents the smoothing parameter, μ B This represents the mean of the output data after processing by the convolutional layer. This represents the variance of the output data after processing by the convolutional layer. γ and β are network training parameters. This represents the new output value after batch normalization layer processing.

[0068] Finally, ReLU was chosen as the activation function, defined as:

[0069] z l =max(y l ,0) (6)

[0070] Among them, z l This represents the output of the convolution unit.

[0071] (2) A stacking operation is performed on the base layers (convolutional units). The left end of the stacked network is connected to a base layer, and the output features at both ends of the network are connected to form a fractal structure with two columns. Based on the self-similarity of fractals, the width and depth of the network can be expanded simultaneously to obtain fractal blocks, such as... Figure 6 The diagram shows the expansion rules for fractal blocks;

[0072]

[0073] Where x represents the input of the fractal block, f c Defined as a sub-block, i.e., a convolutional unit. c represents the number of columns of convolutional units in the fractal block. Indicates a combination operation. This indicates a join operation, which combines multiple inputs into a single output.

[0074] (3) Fractal blocks with different numbers of columns and depths are connected through pooling operations to achieve downsampling and reduce computational complexity. Specifically, this is described as follows:

[0075]

[0076] in, Indicates that it is located in region R ij The value of the element at position (p, q) in the middle. Representing region R ij The max pooling output value of the f-th feature map obtained from the l-th layer.

[0077] (4) A fully connected layer is used to transform the output feature dimension of the last pooling layer of the fractal network into the output dimension. The transformed output is then subjected to a non-linear transformation using the ReLU activation function. Finally, a softmax layer is used to output the probability of each category. The formula is described as follows:

[0078] y i =ReLU(w T x i +b) (9)

[0079]

[0080] Where K represents the number of categories.

[0081] The optimization of fractal networks using a multi-objective differential evolution algorithm during training includes the following steps:

[0082] Step 6.1: Determine the two optimization objectives for the multi-objective differential evolution algorithm to optimize the fractal network;

[0083] This embodiment employs a multi-objective differential evolution algorithm to optimize the fractal network, aiming to find the optimal model with excellent performance and low complexity during training. Therefore, this embodiment determines the time complexity and recognition accuracy of the fractal network as the two objectives of the multi-objective differential evolution algorithm optimization. To simplify computational costs, this embodiment only considers the computational cost of the convolutional layers as the first optimization objective during training. The time complexity of the convolutional layers is defined as follows:

[0084]

[0085] Where B is the number of fractal blocks in the fractal network. M is the number of columns of convolutional units in the b-th block. l is the index of any convolutional layer in the fractal block. l is the number of output channels of the l-th convolutional layer, while M l-1 s is the number of input channels in the l-th convolutional layer. l It is the kernel size in the l-th convolutional layer. It is the output feature size of the l-th convolutional layer.

[0086] Since the two objectives of the multi-objective differential evolution algorithm for optimizing fractal networks are conflicting, this embodiment uses the recognition error rate as the second objective for optimization, defined as:

[0087]

[0088] Where N represents the total number of input samples for the fractal network, and N′ represents the number of correctly identified samples.

[0089] Therefore, the two optimization objectives of the multi-objective differential evolution algorithm for optimizing fractal networks are denoted as:

[0090]

[0091] Here, θ is defined as the parameter to be optimized in the fractal network.

[0092] Step 6.2: Set up a multi-objective differential evolution algorithm to optimize the search space of the fractal network. The search elements include: the number of fractal blocks, the number of columns of convolutional units, the number of channels of convolutional layers, and the learning rate; and determine the hyperparameters in the training process of the fractal network.

[0093] In this embodiment, the range of fractal block count is set to {1, 2, 3, 4}, the range of convolutional unit column count is set to {3, 4, 5}, the range of convolutional layer channel count is set to {8, 16, 32, 64}, and the range of learning rate is set to {0.1, 0.01, 0.001, 0.0001}. The hyperparameter settings during fractal network training include: batch size set to 64, number of iterations set to 150, model optimization strategy set to Adam, loss function selected as CrossEntropy, and dropout value set to 0.1.

[0094] Step 6.3: Determine the relevant parameters of the multi-objective differential evolution algorithm;

[0095] In this embodiment, the population size of the multi-objective differential evolution algorithm is P = 20, meaning it contains 20 individuals. Each individual has a dimension of 4, which is the number of search elements mentioned in step 6.2. The number of generations is G = 5, the mutation factor mu = 0.3, and the crossover probability cr = 0.8. The initial population size is p = 0, and the initial number of generations is g = 0.

[0096] Step 6.4: Based on the search space in Step 3.2, with population size P and individual dimension 4, randomly generate 20 individuals as the initial population;

[0097] Step 6.5: Using the initial population as the parent population, train the fractal network corresponding to each individual in the initial population based on the expanded dataset to obtain two optimization target values ​​for each individual;

[0098] Step 6.6: Perform a difference operation on two random different individuals from the parent population according to the mutation factor mu, and add them to an existing individual to achieve mutation, thereby obtaining a mutated offspring population;

[0099] Step 6.7: Perform crossover operations on each individual in the parent population and each individual in the mutated offspring population according to the crossover probability cr to achieve information exchange, thereby generating a new offspring population. Based on the expanded dataset, train the fractal network corresponding to each individual in the new offspring population to obtain two optimization target values ​​for each individual.

[0100] Step 6.8: Merge the parent population with the new offspring population, and perform a fast non-dominated sorting on the merged population according to the Pareto dominance principle of the multi-objective differential evolution algorithm, prioritizing the placement of individuals with higher rankings into the individual set.

[0101] Step 6.9: Compare the size of the individual set obtained in Step 6.8 with the size of the population size P. If the two are the same, then use the individual set as the parent population of the next generation; otherwise, select individuals from the individual set as the parent population of the next generation according to the crowding distance rule, and increment the generation number g by one.

[0102] Step 6.10: As Figure 7 As shown, repeat steps 6.6 to 6.9 until the number of generations g equals the maximum number of generations G, thus ending the multi-objective differential evolution algorithm optimization process and obtaining the Pareto optimal solution. Each optimal solution corresponds to a fractal network with a different structure.

[0103] Step 7: Determine the Knee point from the Pareto optimal solution obtained in Step 6 using a distance-based method to obtain the metallographic structure identification model;

[0104] This embodiment uses a distance-based method to determine the Knee point, such as... Figure 8 As shown. The specific method is explained below:

[0105] First, connect the two boundary points "A" and "F" of the Pareto front to form a straight line "L", defined as: a1x + a2y + a3 = 0, where the parameters a1, a2, and a3 are determined by the boundary points. Next, calculate the distance from each point on the Pareto front to this straight line; the point with the largest distance is the Knee point. The distance calculation is shown in formula (15):

[0106]

[0107] Step 8: Train the metallographic structure recognition model using the expanded dataset to obtain the trained metallographic structure recognition model;

[0108] In this embodiment, the hyperparameter settings for the training process of the metallographic structure identification model include: batch size set to 64, number of iterations set to 200, learning rate set to 0.008, model optimization strategy selected as Adam, loss function selected as CrossEntropy, and dropout value set to 0.1.

[0109] Step 9: Use the trained metallographic structure recognition model to perform metallographic structure recognition on the metallographic structure image input in the actual application.

[0110] The metallographic data used in this embodiment are derived from publicly available information from the Materials Science Library's Information Technology Promotion Center. This includes numerous metallographic images of materials at various magnifications, each with unique composition and internal structural characteristics. Based on the acquired experimental data, the method of this invention is validated. The fractal dimension values ​​for each metallographic category obtained in this embodiment are shown in Table 1. Texture data is migrated to a self-similar metallographic dataset to enhance the dataset, based on a threshold of 0.1. For example, "blanket-1" is added to the pearlite category, and "lentil-1" is migrated to the austenite category.

[0111] Under the same training parameters and metallographic structure recognition model settings, four different datasets were used to train the metallographic structure recognition model, and the test datasets divided in step 1 were used for testing. Table 2 shows the statistical results of material metallographic structure recognition obtained in this embodiment. The four different datasets are material metallographic structure data (material data), material metallographic structure data expanded based on transfer learning (material + transfer data), material metallographic structure data expanded based on generative adversarial networks (material + generated data), and material metallographic structure data expanded based on transfer learning and generative adversarial networks (material + transfer + generated data). As shown in Table 2, the recognition accuracy reached 69.54% after training with material metallographic structure data. The recognition accuracy reached 83.35% after training with material metallographic structure data expanded based on transfer learning, while the recognition accuracy did not reach a good level when training with material metallographic structure data based on generative adversarial networks, and was lower than that of the material metallographic structure data based on transfer learning. The best recognition result was obtained by training with a mixture of three data sets, with an accuracy rate of 85.49%.

[0112] Table 1 Fractal Dimension

[0113]

[0114] Table 2 Comparison Results of Metallographic Structure Identification of Materials

[0115]

[0116] It should be understood that, inspired by the technical concept of this invention, those skilled in the art can make various improvements or modifications based on the above content without departing from the scope of this invention, and these modifications still fall within the protection scope of this invention.

Claims

1. A method for metallographic structure identification based on a multi-objective fractal evolution network, characterized in that, The method includes the following steps: Collect images of metallographic structures of different categories, establish an original metallographic structure dataset, and divide the dataset into a training dataset and a test dataset according to a preset ratio; Find the fractal dimension D of each metallographic image in the training dataset. B ; Create a fabric texture image dataset and calculate the fractal dimension of each image in the dataset. ; Based on D B and The fabric texture images are transferred to the training dataset, thus augmenting the original training dataset with additional samples; the fractal dimension is then calculated. and The absolute value of the difference , less than the preset threshold The corresponding fabric texture images are transferred to the training dataset to expand the sample data volume of the training dataset; The data from the expanded training dataset is input into the fractal network, and a multi-objective differential evolution algorithm is used to optimize the fractal network during the training process, gradually finding the Pareto optimal solution. The Knee point is determined from the Pareto optimal solution to obtain the metallographic structure identification model; The metallographic structure recognition model was trained using the expanded training dataset to obtain a well-trained metallographic structure recognition model. The trained metallographic structure recognition model is used to perform metallographic structure recognition on the input metallographic structure image.

2. The metallographic structure identification method based on multi-objective fractal evolution network according to claim 1, characterized in that, The fractal dimension D of each metallographic image was calculated using the box counting method. B The fractal dimension of each fabric texture image was calculated using box counting. .

3. The metallographic structure identification method based on multi-objective fractal evolution network according to claim 1, characterized in that, The method also includes the steps of: generating new data that is highly similar to metallographic data using a generative adversarial network based on the original training dataset, adding the new data to the training dataset after the first sample data augmentation, and performing a second sample data augmentation on the training dataset.

4. The metallographic structure identification method based on multi-objective fractal evolution network according to claim 1, characterized in that, With the optimization objectives of minimizing the time complexity and recognition error rate of the fractal network, a Pareto optimal solution that balances these two objectives is gradually found during the optimization of the fractal network using a multi-objective differential evolution algorithm.

5. The metallographic structure identification method based on multi-objective fractal evolution network according to claim 4, characterized in that, Minimizing the time complexity of a fractal network is equivalent to minimizing the time complexity of a convolutional layer.

6. The metallographic structure identification method based on multi-objective fractal evolution network according to claim 1, characterized in that, The distance-based method is used to determine the Knee point from the Pareto optimal solution.

Images