A method for generating CT images of asphalt mixture based on generative adversarial networks

Abstract

A method for generating asphalt mixture CT images based on an adversarial generative network relates to the field of road engineering technology. The existing asphalt mixture CT image generation method does not take into account the aggregate particle spacing, aggregate particle area ratio and void ratio, and cannot truly reflect the actual structural distribution of the asphalt mixture, which leads to the problem of low accuracy of the generated asphalt mixture CT image. The present application introduces asphalt mixture micro-indicators (aggregate particle spacing, aggregate particle area ratio and void ratio) by modifying the loss functions of the generator and the discriminator, thereby improving the similarity between the generated image and the real image, facilitating the detailed processing of the generated image by the generator, and at the same time being more adaptable to the feature distribution of different samples, thereby improving the accuracy of asphalt mixture CT image generation.

Description

Technical Field

[0001] The invention relates to the technical field of road engineering, and in particular to a method for generating asphalt mixture CT images based on an adversarial generative network. Background Art

[0002] Asphalt mixture is a composite material composed of mineral material, asphalt binder and voids. Its composition structure is relatively complex, the test results vary greatly, and the reproducibility is poor. The production process takes a lot of time and materials, and needs to be reproduced continuously to meet the needs of subsequent tests such as testing asphalt performance and improving the production process. In order to reduce costs, improve test efficiency, and ensure the repeatability and controllability of test pieces, the intelligent design of asphalt mixture is an inevitable trend in future development.

[0003] At present, the main research method for the intelligent design of asphalt mixtures is to obtain the internal structure of asphalt mixtures through electronic computed tomography (CT) scanning, conduct volume composition analysis, virtual specimen production, and virtual mechanical tests. This method can reduce the number of asphalt mixture specimens required for the test to a certain extent, but a large number of tests are still required as basic data for research in the early stage, which is costly and inefficient. Therefore, conducting research on asphalt mixture CT image generation is an important cornerstone of intelligent design.

[0004] In recent years, generative adversarial networks, which have the advantages of generative data enhancement, optimization problem solving and creativity, have shown great potential in predicting and optimizing asphalt mixture performance, detecting and identifying defects in asphalt mixture images, etc. Generative adversarial networks are composed of two important parts: the generator and the discriminator. The generator learns the distribution of real samples from real data and generates false samples with similarity close to the real data; the discriminator distinguishes between samples obtained from real data and samples generated by the generator. The two models are iteratively optimized through continuous adversarial training to make the data distribution generated by the generator as close to the real data distribution as possible. Although scholars in the road field have successively applied generative adversarial network technology, there is little research in the field of asphalt mixture CT image generation.

[0005] In real asphalt mixtures, there are usually asphalt mixture microstructures such as aggregate particle spacing, aggregate particle area ratio, and void ratio. However, the existing asphalt mixture CT image generation method does not take into account the aggregate particle spacing, aggregate particle area ratio, and void ratio. Therefore, the asphalt mixture CT image generated by the existing method cannot truly reflect the actual structural distribution of the asphalt mixture, which leads to the problem of low accuracy of the generated asphalt mixture CT image. Summary of the invention

[0006] The purpose of the present invention is to propose a method for generating CT images of asphalt mixtures based on a generative adversarial network, in view of the fact that the existing method for generating CT images of asphalt mixtures does not take into account the aggregate particle spacing, aggregate particle area ratio and void ratio, and cannot truly reflect the actual structural distribution of the asphalt mixture, which leads to the problem that the accuracy of the generated asphalt mixture CT images is low.

[0007] The technical solution adopted by the present invention to solve the above technical problems is:

[0008] A method for generating asphalt mixture CT images based on a generative adversarial network comprises the following steps:

[0009] Step 1: Obtain the original CT image of asphalt mixture;

[0010] Step 2: Convert the original CT image of asphalt mixture into a binary image;

[0011] Step 3: Input the binary image into the trained Wasserstein generative adversarial network to obtain the output asphalt mixture CT image;

[0012] The training process of the Wasserstein generative adversarial network is:

[0013] Step 31: using step 1 and step 2 to obtain a binary image, the binary image includes a black part and a white part, and then extracting features of the black part in the binary image to obtain a gap feature image;

[0014] Step 32: Set a particle size threshold, and divide the white part in the binary image into a white part not less than the particle size threshold and a white part less than the particle size threshold according to the particle size threshold;

[0015] Step 33: extract features of the white part that is not less than the particle size threshold to obtain an aggregate feature image;

[0016] Step 3 and 4: Extract features of the white part that is smaller than the particle size threshold to obtain a mortar feature image;

[0017] Step 35: annotate the void feature image, aggregate feature image and mortar feature image respectively, and merge the annotated void feature image, aggregate feature image and mortar feature image to obtain the real CT image of asphalt mixture, so as to construct the training set;

[0018] Step 36: Use the training set to train the Wasserstein generative adversarial network to obtain a trained Wasserstein generative adversarial network;

[0019] The loss function of the discriminator in the Wasserstein generative adversarial network is:

[0020]

[0021] Among them, L, L D are the total loss function of the discriminator and the loss function of the Wasserstein generative adversarial network, D(x i ), D(G(z i )) are the output of the discriminator for real samples and the output of the discriminator for generated samples, respectively. P , m Q are the number of samples of real images and generated images, respectively, HI (P, Q), MSE1, MSE2 are the cross distance loss of aggregate particle spacing, the mean square error loss of aggregate particle area ratio, and the mean square error loss of void ratio, respectively. λ1, λ2, λ3 are the weights corresponding to the additional loss function, i is the i-th sample,

[0022] The loss function of the generator in the Wasserstein generative adversarial network is:

[0023]

[0024] Among them, D(G(z i )) is the output of the generator for the generated sample, m Q is the number of samples to generate the image.

[0025] Furthermore, the specific steps of step one are:

[0026] Step 1: Prepare asphalt mixture and make it into Marshall specimen;

[0027] Step 1 and 2: Scan the Marshall specimen to obtain the original CT image of the asphalt mixture specimen. The size of the original CT image of the asphalt mixture specimen is 800×800.

[0028] Furthermore, the specific steps of step 2 are:

[0029] Step 21: Filter and reduce noise and enhance grayscale of the original CT image of asphalt mixture;

[0030] Step 22: For the image after grayscale enhancement in step 21, use the circular Otsu algorithm to perform threshold segmentation to obtain a binary image.

[0031] Furthermore, the particle size threshold is 4.5 mm.

[0032] Furthermore, the generator includes 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a ReLU activation function;

[0033] The discriminator includes 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a LeakyReLU activation function.

[0034] Furthermore, the generator is expressed as:

[0035]

[0036] Among them, x i+m，j+n ConvTranspose2d(x) is the input tensor of row i+n and column j+m. i，j is the output tensor of the i-th row and j-th column, m and n are the indices of the deconvolution kernel weights, m and n represent the rows and columns in the weight matrix respectively, M and N represent the total number of rows and columns in the weight matrix respectively, and K m,n is the weight value of the convolution kernel in the mth row and nth column.

[0037] Furthermore, the discriminator is expressed as:

[0038]

[0039]

[0040] Among them, H in , W in is the length and width of the input image, H out , W out is the length and width of the output image, S is the stride, P is the padding, and K is the convolution kernel weight.

[0041] Furthermore, the D HI (P, Q) is expressed as:

[0042]

[0043] Among them, P and Q are the real image and the generated image respectively, p i ,q i are the normalized frequencies of the intervals of the discretized continuous distribution for the real image and the generated image, respectively.

[0044] Further, the MSE1 is expressed as:

[0045]

[0046]

[0047] in, are the real image and generated aggregate particle area ratio, are the number of pixels in all particle areas and the number of pixels in the entire image, respectively.

[0048] Further, the MSE2 is expressed as:

[0049]

[0050]

[0051] in, are the probabilities of the i-th void fraction interval in the real image and the generated image, respectively. are the void area and the area of ​​the entire mixture cross section respectively.

[0052] The beneficial effects of the present invention are:

[0053] This application introduces asphalt mixture micro-indicators (aggregate particle spacing, aggregate particle area ratio, and void ratio) by modifying the loss functions of the generator and the discriminator, thereby improving the similarity between the generated image and the real image, facilitating the generator to process the details of the generated image, and being more adaptable to the feature distribution of different samples, thereby improving the accuracy of asphalt mixture CT image generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is the original image of the asphalt mixture CT part;

[0055] Figure 2 Schematic diagram for image feature extraction Figure 1 ;

[0056] Figure 3 Schematic diagram for image feature extraction Figure 2 ;

[0057] Figure 4 Schematic diagram for image feature extraction Figure 3 ;

[0058] Figure 5 Schematic diagram of training samples;

[0059] Figure 6 This is a schematic diagram of the network structure of this application;

[0060] Figure 7 Schematic diagram of image generation. DETAILED DESCRIPTION

[0061] It should be particularly noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0062] Specific implementation method 1: refer to Figure 1 Specifically describing this embodiment, a method for generating a CT image of an asphalt mixture based on a generative adversarial network described in this embodiment includes the following steps:

[0063] Step 1: Obtain the original CT image of asphalt mixture;

[0064] Step 2: Convert the original CT image of asphalt mixture into a binary image;

[0065] Step 3: Input the binary image into the trained Wasserstein generative adversarial network to obtain the output asphalt mixture CT image;

[0066] This application constructs a generative adversarial network model that integrates asphalt mixture meso-indicators, which mainly includes: using the Wasserstein generative adversarial network as the model framework, introducing asphalt mixture meso-indicators into the loss functions of the generator and the discriminator, and training the model.

[0067] The training process of the Wasserstein generative adversarial network is:

[0068] Step 31: using step 1 and step 2 to obtain a binary image, the binary image includes a black part and a white part, and then extracting features of the black part in the binary image to obtain a gap feature image;

[0069] Step 32: Set a particle size threshold, and divide the white part in the binary image into a white part not less than the particle size threshold and a white part less than the particle size threshold according to the particle size threshold;

[0070] Step 33: extract features of the white part that is not less than the particle size threshold to obtain an aggregate feature image;

[0071] Step 3 and 4: Extract features of the white part that is smaller than the particle size threshold to obtain a mortar feature image;

[0072] Step 35: annotate the void feature image, aggregate feature image and mortar feature image respectively, and merge the annotated void feature image, aggregate feature image and mortar feature image to obtain the real CT image of asphalt mixture, so as to construct the training set;

[0073] Step 36: Use the training set to train the Wasserstein generative adversarial network to obtain a trained Wasserstein generative adversarial network;

[0074] The loss function of the discriminator in the Wasserstein generative adversarial network is:

[0075]

[0076] Among them, L, L D are the total loss function of the discriminator and the loss function of the Wasserstein generative adversarial network, D(x i), D(G(z i )) are the output of the discriminator for real samples and the output of the discriminator for generated samples, respectively. P , m Q are the number of samples of real images and generated images, respectively, HI (P, Q), MSE1, MSE2 are the cross distance loss of aggregate particle spacing, the mean square error loss of aggregate particle area ratio, and the mean square error loss of void ratio, respectively. λ1, λ2, λ3 are the weights corresponding to the additional loss function, i is the i-th sample,

[0077] The loss function of the generator in the Wasserstein generative adversarial network is:

[0078]

[0079] Among them, D(G(z i )) is the output of the generator for the generated sample, m Q is the number of samples to generate the image.

[0080] This application designs a generative adversarial network model that integrates asphalt mixture micro-indicators, such as Figure 6 , mainly including: using Wasserstein generative adversarial network as the model framework, introducing asphalt mixture micro-indicators into the loss functions of the generator and discriminator, and training the model, specifically:

[0081] Based on the Wasserstein generative adversarial network, the encoder of the generator is set, including 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a ReLU rectified linear unit activation function to extract the aggregate distribution and gap distribution features in the image;

[0082] Improve the Wasserstein generative adversarial network and set the encoder of the discriminator, including 8 network layers: each network layer includes 1 4×4 2D convolution layer, 1 normalization layer and 1 LeakyReLU with leaky rectified linear unit activation function;

[0083] After that, the total loss function of the discriminator is designed. On the basis of the loss function of the Wasserstein generative adversarial network, the evaluation indicators of aggregate distribution and void distribution in the image are added, including the histogram intersection distance loss of aggregate particle spacing, the mean square error loss of aggregate particle area ratio, and the mean square error loss of void ratio. Specifically, they include:

[0084] ① Cross-distance loss of aggregate particle spacing: Calculate the cross-distance between the real image and the generated image aggregate particle spacing;

[0085] ② Mean square error loss of aggregate particle area ratio: Calculate the mean square error of the aggregate particle area ratio between the real image and the generated image;

[0086] ③ Mean square error loss of void ratio: Calculate the mean square error of the aggregate particle area ratio between the real image and the generated image;

[0087] ④ Discriminator total loss function: Based on the loss function of the Wasserstein generative adversarial network, the cross distance loss of aggregate particle spacing, the mean square error loss of aggregate particle area ratio, and the mean square error loss of void ratio are added as additional losses to the loss function of the discriminator of the generative adversarial network, and the weights are set in turn.

[0088] Finally, the total loss function of the generator is designed: the loss function of the generator is the negative average output of the generated samples on the discriminator.

[0089] The Adam optimizer with gradient penalty in the Wasserstein generative adversarial network is used to optimize the parameters of the generator and discriminator respectively.

[0090] In order to fill the existing gaps, improve test efficiency, and promote the development of intelligent design of asphalt mixtures, this application proposes a method for generating CT images of asphalt mixtures based on adversarial generative networks. This method is an important cornerstone of intelligent design of asphalt mixtures and can be used as a data source for its intelligent design. It can greatly enrich the asphalt mixture CT image database and reduce the preliminary preparations for researchers to study asphalt mixtures.

[0091] This application improves the generalization ability of the model and solves the problem that in the process of using a generative adversarial network to generate asphalt mixture CT images, the generated images are similar to real images only in morphology, but do not meet or meet the requirements for the use of asphalt mixture engineering in terms of microstructure.

[0092] This application can better generate CT images of asphalt mixtures, effectively solve the massive data sets required for the intelligent design of asphalt mixtures, reduce the cost and time of laboratory testing, promote the process of intelligent design of asphalt mixtures, and assist the field of smart materials.

[0093] Specific implementation method 2: This implementation method is a further description of the specific implementation method 1. The difference between this implementation method and the specific implementation method 1 is that the specific steps of step 1 are:

[0094] Step 1: Prepare asphalt mixture and make it into Marshall specimen;

[0095] Step 1 and 2: Scan the Marshall specimen to obtain the original CT image of the asphalt mixture specimen. The size of the original CT image of the asphalt mixture specimen is 800×800.

[0096] Specific implementation method three: This implementation method is a further description of specific implementation method two. The difference between this implementation method and specific implementation method two is that the specific steps of step two are:

[0097] Step 21: Filter and reduce noise and enhance grayscale of the original CT image of asphalt mixture;

[0098] Step 22: For the image after grayscale enhancement in step 21, use the circular Otsu algorithm to perform threshold segmentation to obtain a binary image.

[0099] The present application performs filtering, noise reduction and grayscale enhancement on the acquired CT image, and then uses the circular Otsu algorithm for threshold segmentation; converts the image into a black-and-white binary image with only voids, where the voids are black, extracts the void feature image, and identifies the white parts as aggregates and mortar; extracts the aggregate feature image, screens out the black aggregates with a particle size of less than 4.5 mm in the image, and treats them as part of the mortar; extracts the mortar feature image; color-labels the extracted aggregate, void, and mortar feature images, and synthesizes a labeled real CT image of asphalt mixture; divides the data set of the real CT image of asphalt mixture after preprocessing into a training set, a test set, and a validation set.

[0100] Specific implementation method 4: This implementation method is a further description of specific implementation method 3. The difference between this implementation method and specific implementation method 3 is that the particle size threshold is greater than 4.5 mm.

[0101] Specific implementation mode five: This implementation mode is a further explanation of specific implementation mode four. The difference between this implementation mode and specific implementation mode four is that the generator includes 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a ReLU activation function;

[0102] The discriminator includes 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a LeakyReLU activation function.

[0103] Specific implementation method 6: This implementation method is a further explanation of the specific implementation method 5. The difference between this implementation method and the specific implementation method 5 is that the generator is represented as:

[0104]

[0105] Among them, x i+m，j+n ConvTranspose2d(x) is the input tensor of row i+n and column j+m. i，jis the output tensor of the i-th row and j-th column, m and n are the indices of the deconvolution kernel weights, m and n represent the rows and columns in the weight matrix respectively, M and N represent the total number of rows and columns in the weight matrix respectively, and K m，n is the weight value of the convolution kernel in the mth row and nth column.

[0106] Specific implementation method 7: This implementation method is a further explanation of specific implementation method 6. The difference between this implementation method and specific implementation method 6 is that the discriminator is represented as:

[0107]

[0108]

[0109] Among them, H in , W in is the length and width of the input image, H out , W out is the length and width of the output image, S is the stride, P is the padding, and K is the convolution kernel weight.

[0110] Specific implementation method 8: This implementation method is a further description of specific implementation method 7. The difference between this implementation method and specific implementation method 7 is that the D HI (P, Q) is expressed as:

[0111]

[0112] Among them, P and Q are the real image and the generated image respectively, p i ,q i are the normalized frequencies of the intervals of the discretized continuous distribution for the real image and the generated image, respectively.

[0113] Specific implementation method 9: This implementation method is a further description of specific implementation method 8. The difference between this implementation method and specific implementation method 8 is that the MSE1 is represented as:

[0114]

[0115]

[0116] in, are the real image and generated aggregate particle area ratio, are the number of pixels in all particle areas and the number of pixels in the entire image, respectively.

[0117] Specific implementation method ten: This implementation method is a further description of specific implementation method nine. The difference between this implementation method and specific implementation method nine is that the MSE2 is represented as:

[0118]

[0119]

[0120] in, are the probabilities of the i-th void fraction interval in the real image and the generated image, respectively. are the void area and the area of ​​the entire mixture cross section respectively.

[0121] Example:

[0122] Step 1: Prepare asphalt mixture and make it into a standard Marshall specimen with a diameter of (101.60.2) mm. Use the micro-focus industrial CT of Phoenix Company in Germany to scan the specimen and obtain the original CT image of the asphalt mixture specimen with a size of 800×800. Figure 1 .

[0123] Step 2: Filter and reduce noise and enhance grayscale of the acquired CT image, such as Figure 2 , and then use the circular Otsu algorithm for threshold segmentation; convert the image into a black and white binary image with only gaps, the gaps are black, and extract the gap feature image, such as Figure 3 , the white part is identified as aggregate and mortar; extract the aggregate feature image, the aggregate is white, screen out the white aggregate with a particle size of less than 4.5 mm in the image, and regard it as part of the mortar; extract the mortar feature image, such as Figure 4 , white is mortar; the extracted aggregate, void, and mortar feature images are color-labeled and synthesized into a labeled real CT image of asphalt mixture, such as Figure 5 ; The real CT image data set of asphalt mixture after preprocessing is divided into training set, test set and validation set.

[0124] Step 3: Based on the Wasserstein generative adversarial network, set the encoder of the generator, including 8 network layers, each of which includes 1 4×4 two-dimensional deconvolution layer, 1 normalization layer and 1 ReLU rectified linear unit activation function, which is used to extract the aggregate distribution and void distribution characteristics in the image; set the encoder of the discriminator, including 8 network layers: each network layer includes 1 4×4 two-dimensional convolution layer, 1 normalization layer and 1 LeakyReLU rectified linear unit activation function with leakage; design the total loss function of the discriminator, and add the histogram cross distance loss of the aggregate particle spacing, the mean square error loss of the aggregate particle area ratio, and the mean square error loss of the void ratio to the loss function of the Wasserstein generative adversarial network; use the gradient penalty Adam optimizer to optimize the parameters of the generator and discriminator respectively.

[0125] Step 4: Input the asphalt mixture CT image into the trained model to generate the asphalt mixture CT image, such as Figure 7 .

[0126] It should be noted that the specific implementation is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the scope of protection of the rights. Any partial changes made according to the claims and description of the present invention should still fall within the scope of protection of the present invention.

Claims

1. A method for generating CT images of asphalt mixture based on a generative adversarial network, characterized in that The following steps are involved: Step 1: Obtain the original CT image of asphalt mixture; Step 2: Convert the original CT image of asphalt mixture into a binary image; Step 3: Input the binary image into the trained Wasserstein generative adversarial network to obtain the output asphalt mixture CT image; The training process of the Wasserstein generative adversarial network is: Step 31: using step 1 and step 2 to obtain a binary image, the binary image includes a black part and a white part, and then extracting features of the black part in the binary image to obtain a gap feature image; Step 32: Set a particle size threshold, and divide the white part in the binary image into a white part not less than the particle size threshold and a white part less than the particle size threshold according to the particle size threshold; Step 33: extract features of the white part that is not less than the particle size threshold to obtain an aggregate feature image; Step 3 and 4: Extract features of the white part that is smaller than the particle size threshold to obtain a mortar feature image; Step 35: annotate the void feature image, aggregate feature image and mortar feature image respectively, and merge the annotated void feature image, aggregate feature image and mortar feature image to obtain the real CT image of asphalt mixture, so as to construct the training set; Step 36: Use the training set to train the Wasserstein generative adversarial network to obtain a trained Wasserstein generative adversarial network; The loss function of the discriminator in the Wasserstein generative adversarial network is: Among them, L D , L are the total loss function of the discriminator and the loss function of the Wasserstein generative adversarial network, D(x i ), D(G(z i )) are the output of the discriminator for real samples and the output of the discriminator for generated samples, respectively. P , m Q are the number of samples of real images and generated images, respectively, HI (P, Q), MSE1, MSE2 are the cross distance loss of aggregate particle spacing, the mean square error loss of aggregate particle area ratio, and the mean square error loss of void ratio, respectively. λ1, λ2, λ3 are the weights corresponding to the additional loss function, i is the i-th sample, The loss function of the generator in the Wasserstein generative adversarial network is: Among them, D(G(z i )) is the output of the generator for the generated sample, m Q is the number of samples to generate the image.

2. The asphalt mixture CT image generation method based on adversarial generative network according to claim 1 is characterized in that The specific steps of step one are: Step 1: Prepare asphalt mixture and make it into Marshall specimen; Step 1 and 2: Scan the Marshall specimen to obtain the original CT image of the asphalt mixture specimen. The size of the original CT image of the asphalt mixture specimen is 800×800.

3. The asphalt mixture CT image generation method based on adversarial generative network according to claim 2 is characterized in that The specific steps of step 2 are: Step 21: Filter and reduce noise and enhance grayscale of the original CT image of asphalt mixture; Step 22: For the image after grayscale enhancement in step 21, use the circular Otsu algorithm to perform threshold segmentation to obtain a binary image.

4. The asphalt mixture CT image generation method based on adversarial generative network according to claim 2 is characterized in that The particle size threshold is 4.5 mm.

5. The asphalt mixture CT image generation method based on adversarial generative network according to claim 2 is characterized in that The generator includes 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a ReLU activation function; The discriminator includes 8 network layers, each of which includes a 4×4 two-dimensional deconvolution layer, a normalization layer and a LeakyReLU activation function.

6. The asphalt mixture CT image generation method based on adversarial generative network according to claim 5 is characterized in that The generator is expressed as: Among them, x i+m，j+n ConvTranspos2d(x) is the input tensor of row i+n and column j+m. i，j is the output tensor of the i-th row and j-th column, m and n are the indices of the deconvolution kernel weights, m and n represent the rows and columns in the weight matrix respectively, M and N represent the total number of rows and columns in the weight matrix respectively, and K m，n is the weight value of the convolution kernel in the mth row and nth column.

7. The asphalt mixture CT image generation method based on adversarial generative network according to claim 6 is characterized in that The discriminator is expressed as: Among them, H in , W in is the length and width of the input image, H out , W out is the length and width of the output image, S is the stride, P is the padding, and K is the convolution kernel weight.

8. The asphalt mixture CT image generation method based on adversarial generative network according to claim 7 is characterized in that The D HI (P, Q) is expressed as: Among them, P and Q are the real image and the generated image respectively, p i ,q i are the normalized frequencies of the intervals of the discretized continuous distribution for the real image and the generated image, respectively.

9. The asphalt mixture CT image generation method based on adversarial generative network according to claim 8 is characterized in that The MSE1 is expressed as: in, are the real image and generated aggregate particle area ratio, are the number of pixels in all particle areas and the number of pixels in the entire image, respectively.

10. The asphalt mixture CT image generation method based on adversarial generative network according to claim 9, characterized in that The MSE2 is expressed as: in, are the probabilities of the i-th void fraction interval in the real image and the generated image, respectively. are the void area and the area of ​​the entire mixture cross section respectively.

Images