Asphalt mixture slice map migration method and device based on improved image translation neural network, and medium

Abstract

The invention relates to an asphalt mixture slice map migration method and device based on an improved image translation neural network, and a medium. The method comprises the following steps: S1, obtaining part of real asphalt mixture slice maps of different gradation and aggregate; s2, cutting the obtained real slice image, and constructing an asphalt mixture data set based on the cut real slice image; s3, a training set and a test set are constructed based on the asphalt mixture data set, a corresponding migration model is trained based on the training set, and the migration model is constructed based on an SIT-GAN network; s4, testing the trained migration model by using the test set; and S5, based on the aggregate and gradation of the input picture, combining the aggregate of the picture to be migrated and selecting a corresponding migration model, and inputting the input picture into the migration model to obtain a migrated asphalt mixture slice map output by the migration model. Compared with the prior art, the method has the advantages that the perceptual capacity of fine aggregate, mortar, pores and other details is increased, the perceptual and extraction capacity of the migration model for complex multi-scale features of the asphalt mixture is improved, and the quality of the generated slice map is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of road engineering, and in particular to an asphalt mixture slice image migration method, device and medium based on an improved image translation neural network. Background Art

[0002] With the rapid development of image processing technology, its broad application prospects in the field of road engineering are gradually emerging. Asphalt slice images are the most intuitive image information that can reflect the various characteristics of asphalt mixtures. They play a key role in road material analysis, quality control, performance prediction and engineering design. Asphalt mixture slice images can intuitively display the structure of the mixture, including the distribution of aggregates, asphalt binders and pores. At the same time, the slice image analysis of the samples after construction can detect the gradation, evaluate the compaction effect, the uniformity of the mixture and possible construction defects. In addition, slice images are helpful in studying the formation and expansion of cracks, the structural changes of asphalt after aging, etc. At present, there are many cases and applications of using threshold segmentation, convolutional neural networks, multi-scale fusion, autoencoders, time series analysis networks and other technologies to process slice images, thereby completing related research cases and applications.

[0003] A major feature of these technologies is that they have a huge demand for data volume. Therefore, in order to complete related research, a large dataset of asphalt mixture slice images is required. However, in practical applications, it is costly to obtain asphalt slice images. From the perspective of monetary cost, making a large number of asphalt mixture specimens with different gradations, different asphalts, and different porosities, and then taking slice images, requires a series of equipment and material expenses, which obviously requires a high monetary cost; from the perspective of time cost, the whole process from the preparation of specimens to the shooting of a large number of slice images is very time-consuming, and also requires a lot of manpower experimental operations. If the specimens are drilled directly from the road surface, the integrity of the road surface will be destroyed. In short, the current method of relying entirely on mixture specimens to obtain slice images has many disadvantages and shortcomings. Summary of the invention

[0004] The purpose of the present invention is to provide a method, device and medium for migrating asphalt mixture slice images based on an improved image translation neural network.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] An asphalt mixture slice image migration method based on an improved image translation neural network, comprising:

[0007] Step S1: obtaining real slice images of asphalt mixtures with different gradations and aggregates;

[0008] Step S2: cropping the acquired real slice image, and constructing an asphalt mixture dataset based on the cropped real slice image;

[0009] Step S3: constructing a training set and a test set based on the asphalt mixture data set, and training a corresponding migration model based on the training set, wherein the migration model is constructed based on the SIT-GAN network;

[0010] Step S4: Use the test set to test the trained migration model;

[0011] Step S5: Based on the aggregate and gradation of the input image, the corresponding migration model is selected in combination with the aggregate and gradation of the image to be migrated, and the input image is input into the migration model to obtain the migrated asphalt mixture slice image output by the migration model.

[0012] The step S1 comprises:

[0013] Photos of all surfaces of a plurality of asphalt mixture specimens are obtained as real slice images, wherein the asphalt mixture specimens are polyhedrons, and at least one of the aggregate and the gradation of any two asphalt mixture specimens is different.

[0014] The asphalt mixture specimen is a rectangular parallelepiped.

[0015] The step S2 comprises:

[0016] Step S2-1: removing invalid parts from all real slice images by clipping, and taking all clipped real slice images as original data sets, wherein the invalid parts at least include background areas, damaged areas, crack areas, reflective areas and stain areas, and the clipped real slice images are rectangular;

[0017] Step S2-2: for all the real slice images in the original data set, each real slice image is cropped into at least one sub-slice image according to the aspect ratio of each real slice image, wherein the sub-slice image is a square, and the side length of any sub-slice image is consistent with the width of the corresponding real slice image;

[0018] Step S2-3: All sub-slice images are combined with their aggregate and gradation information to form an element of the asphalt mixture data set, thereby obtaining the asphalt mixture data set.

[0019] Among all the sub-slice images obtained by cropping the real slice images in the original data set, the similarity between any two sub-slice images is lower than a preconfigured first similarity threshold.

[0020] The migration model includes a generator and a discriminator, and the generator includes:

[0021] 1 reflection filling layer, used to extract and obtain the first feature map fm1 based on the sub-slice map;

[0022] 1 convolution layer, 10 HFEM A modules, 1 activation function layer, 1 convolution layer, 10 HFEM B modules, 1 activation function layer and 1 convolution layer, 10 HFEM C modules, 1 activation function layer connected in sequence, used to extract and generate a second feature map fm2 based on the first feature map fm1;

[0023] 3 layers of transposed convolution layers and activation function layers are used to further expand the size of the second feature map fm2 and compress the dimension to obtain the third feature map fm3;

[0024] A reflection padding layer and a second 7×7 convolutional layer connected in sequence are used to obtain a migrated image based on the third feature map fm3.

[0025] The step S3 comprises:

[0026] Step S3-1: transform the size of all sub-slice images in the asphalt mixture dataset into input suitable for the migration model;

[0027] Step S3-2: constructing a training set and a test set to construct a sample set based on the asphalt mixture data set, and dividing the sample set into a training set and a test set;

[0028] Step S3-3: selecting a corresponding migration model according to the aggregate and gradation information of two sub-slice images of the training set, wherein at least one of the aggregate and gradation of the two sub-slice images is different;

[0029] Step S3-4: Input the first sub-slice image to the generator F of the first migration model A→B , obtained by the generator F A→B Generated fake B-image B , and the generator F that feeds the second sub-slice graph into the second migration model B→A , obtained by the generator F B→A Generated fake A image fake A , where A is the domain of the first sub-slice map, representing the aggregate and gradation information of the first sub-slice map, and B is the domain of the second sub-slice map, representing the aggregate and gradation information of the second sub-slice map;

[0030] Step S3-5: Combine the second sub-slice image and the obtained fake B image B Input to the discriminator D of the first migration model B , respectively get the real domain B image fake B The discrimination result of the fake domain B image and the discrimination result of the fake domain A image are obtained by combining the first sub-slice image with the fake domain A image fakeA Input to the discriminator D of the second migration model A , respectively obtain the discrimination results of the real domain A image and the fake domain A image fake A The judgment result of

[0031] Step S3-6: Based on the discrimination result of the real domain B image and the fake domain B image fake B The first adversarial loss is calculated based on the discrimination results of the real domain A image and the fake domain A image fake A The discrimination result calculates the second adversarial loss;

[0032] Step S3-7: Generate fake B image fake B Input to the generator F of the second migration model B→A , and obtain the regenerated image regen of domain A A , the generated fake A image fake A Input to the generator F of the first migration model A→B , and obtain the regenerated image regen of domain B B , and calculate the cycle consistency loss;

[0033] Step S3-8: Calculate the total loss based on the first adversarial loss, the second adversarial loss and the cycle consistency loss function, and correct the parameters of the migration model based on the total loss.

[0034] The step S4 comprises:

[0035] Step S4-1: Input the real domain A image in the test sample into the generator F A→B Get the fake domain B image, and input the real domain B image in the test sample into the generator F B→A Get the fake domain A image,

[0036] Step S4-2: Calculate the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the fake domain A image and the true domain A image respectively, and calculate the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the true domain B and the fake domain B images respectively;

[0037] Step S4-3: Based on the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the fake domain A image and the true domain A image, combined with the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the true domain B and the fake domain B images, determine whether the migration model is trained.

[0038] An asphalt mixture slice map migration device based on an improved image translation neural network includes a memory, a processor, and a program stored in the memory. When the processor executes the program, the method as described above is implemented.

[0039] A storage medium stores a program, which implements the above method when executed.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. A method for asphalt mixture slice image migration based on an improved image translation neural network is proposed, which realizes the generation of multiple types of slice images in large quantities and with high quality from one type of asphalt mixture slice image.

[0042] 2. When cropping, ensure that the side length is the width of the original rectangle, so that more slices can be generated efficiently while ensuring the integrity of information in one dimension and avoiding the problem of inconsistent scale caused by scaling.

[0043] 3. By using a multi-branch convolutional structure, introducing asymmetric convolution and deformable convolution, the HFEM module is proposed, and used to improve the generator of CycleGAN. The SIT-GAN is innovatively proposed, so that the migration model can improve the perception of details such as fine aggregate, mortar, pores, etc., thereby greatly improving the quality and realism of the generated slice images.

[0044] 4. Using the trained SIT-GAN network, multiple slice images can be generated in large quantities and efficiently when obtaining one asphalt mixture slice image, which greatly reduces the cost of obtaining slice images. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the technical route of the present invention;

[0046] Figure 2 The upper part is the generator structure diagram of SIT-GAN, and the lower part is the HFEM A / B / C structure diagram;

[0047] Figure 3 It is a schematic flow chart of the main steps of the method of the present invention;

[0048] Figure 4 Figure 2 shows the migration effect of asphalt mixture slices, where (a) is AC13 basalt and SMA13 basalt, (b) is AC13 basalt and AC10 basalt, (c) is AC13 basalt and RAC13 limestone, (d) is A13 basalt and AC10 basalt, (e) is SMA13 basalt and RAC13 limestone, and (f) is AC10 basalt and RAC13 limestone. DETAILED DESCRIPTION

[0049] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0050] A method for migrating asphalt mixture slice images based on improved image translation neural network, such as Figure 1 and Figure 3 As shown, including:

[0051] Step S1: Obtaining some real slices of asphalt mixtures with different gradations and aggregates, including:

[0052] Take photos of all surfaces of multiple asphalt mixture specimens as real slice images, where the asphalt mixture specimen is a polyhedron and the aggregates (denoted as M) of any two asphalt mixture specimens are j ) and gradation (denoted as G i ) are different.

[0053] In this embodiment, the asphalt mixture specimen is a rectangular parallelepiped. Specifically, a piece of A3 printing paper is spread flat on a tray, and the specimen is placed horizontally on the paper. A high-definition camera and a tripod are used to photograph each surface of the specimen from a bird's-eye view, and the gradation and aggregate material of each specimen are recorded.

[0054] Step S2: cropping the acquired real slice image, and constructing an asphalt mixture dataset based on the cropped real slice image, including:

[0055] Step S2-1: Remove the invalid parts of all real slice images by clipping, and use all the clipped real slice images as the original data set P = {Pic1, Pic2, ..., Pic n}, wherein the invalid part at least includes the background area, the damaged area, the crack area, the reflective area and the stain area, and the true slice image after clipping is a rectangle;

[0056] Step S2-2: For all the real slices in the original data set, according to the aspect ratio of each real slice Each real slice image is cut into at least one sub-slice image, wherein the sub-slice image is a square, and the side length of any sub-slice image is consistent with the width of its corresponding real slice image;

[0057] This process can be implemented using the OpenCV image processing library. k Cut into m k A square sub-image Where m k =f(r k ), f(r k) needs to balance the number and similarity of the sub-graphs to be cut, that is, to make the number of sub-graphs as large as possible while ensuring the similarity between sub-graphs is low.

[0058] Among all the sub-slice images cropped from the real slice images in the original data set, the similarity between any two sub-slice images is lower than a pre-configured first similarity threshold.

[0059] Step S2-3: All sub-slice images, combined with their aggregate and gradation information, form an element in the asphalt mixture dataset to obtain the asphalt mixture dataset Data. G,M :

[0060]

[0061] Step S3: construct a training set and a test set based on the asphalt mixture data set, and train the corresponding migration model based on the training set. The migration model is constructed based on the CycleGAN network, but improved. The improved model is the SIT-GAN network, as follows:

[0062] The migration model includes a generator and a discriminator, such as Figure 2 As shown, the generator includes:

[0063] 1 reflection filling layer, used to extract the first feature map fm1=(302,302,3) based on the sub-slice map;

[0064] 1 convolution layer, 10 HFEM A modules, 1 activation function layer, 1 convolution layer, 10 HFEM B modules, 1 activation function layer and 1 convolution layer, 10 HFEM C modules, 1 activation function layer connected in sequence, used to extract and generate a second feature map fm2=(64,256,256) based on the first feature map fm1;

[0065] 3 transposed convolution layers and activation function layers are used to further expand the size of the second feature map fm2 and compress the dimension to obtain the third feature map fm3 = (256, 256, 64);

[0066] A reflection filling layer and a second 7×7 convolutional layer are connected in sequence to obtain a migrated image based on the third feature map fm3. The specific size is 256×256×3 (number of channels), which is the same as the size of the input image.

[0067] The HFEM A / B / C modules are composed of three branches, each of which is composed of 1, 2, or 3 convolutions or deformable convolutions in series. For the HFEM A module, the second branch uses 3×3 deformable convolutions, and the third branch uses 3×3 convolutions; for the HFEM B module, the second branch uses 7×7 deformable convolutions, and the third branch uses 7×1 and 1×7 asymmetric convolutions; for the HFEM C module, the second branch uses 3×3 deformable convolutions, and the third branch uses 3×1 and 1×3 asymmetric convolutions. The results of the convolution of each branch are concatenated, then passed through a layer of 1×1 convolution layer, and the output is calculated by weighted connection with the input.

[0068] This application aims to solve the problem of high cost and difficulty in obtaining slice images of asphalt mixtures. Based on the image migration model, this application realizes the mutual migration of different types of slice images, effectively improving the data acquisition efficiency of mixture slice images. At the same time, based on the original CycleGAN model, this application introduces a self-designed multi-scale feature extraction module, adopts a multi-branch convolution splicing structure, and adds asymmetric convolution and deformable convolution. The above improvements enhance the migration model's ability to perceive and extract the complex multi-scale features of asphalt mixtures, and improve the quality of generated slice images.

[0069] Step S3 includes:

[0070] Step S3-1: transform the size of all sub-slice images in the asphalt mixture dataset into an input suitable for the migration model, that is, a size of 256×256 and a channel number of 3;

[0071] Step S3-2: constructing a training set and a test set to construct a sample set based on the asphalt mixture data set, and dividing the sample set into a training set and a test set;

[0072] Step S3-3: selecting a corresponding migration model according to the aggregate and gradation information of two sub-slice images of the training set, wherein at least one of the aggregate and gradation of the two sub-slice images is different;

[0073] Step S3-4: Input the first sub-slice image to the generator F of the first migration model A→B , obtained by the generator F A→B Generated fake B-image B , and the generator F that feeds the second sub-slice graph into the second migration model B→A , obtained by the generator F B→A Generated fake A image fake A , where A is the domain of the first sub-slice map, representing the aggregate and gradation information of the first sub-slice map, and B is the domain of the second sub-slice map, representing the aggregate and gradation information of the second sub-slice map;

[0074] Step S3-5: Combine the second sub-slice image and the obtained fake B image B Input to the discriminator D of the first migration model B , respectively get the real domain B image fake B The discrimination result of the fake domain B image and the discrimination result of the fake domain A image are obtained by combining the first sub-slice image with the fake domain A image fake A Input to the discriminator D of the second migration model A , respectively obtain the discrimination results of the real domain A image and the fake domain A image fake A The judgment result of

[0075] Step S3-6: Based on the discrimination result of the real domain B image and the fake domain B image fake B The discriminant result calculates the first adversarial loss:

[0076]

[0077] Among them: Loss GAN (F A→B ,D B ,A,B) is the first adversarial loss, D B (real B ) is the discriminant result of the real domain B image, D B (fake B ) is a fake domain B image fake B The judgment result of is the mathematical expectation that the true B image follows the distribution of the domain B image, is the mathematical expectation that the true A image follows the distribution of the domain A image,

[0078] And based on the discrimination results of the real domain A image and the fake domain A image fake A The discriminant results calculate the second adversarial loss:

[0079]

[0080] Among them: Loss GAN (F B→A ,D A ,B,A) is the second adversarial loss, D A (real A ) is the discrimination result of the real domain A image, D A (fake A ) is a fake domain A image fake A The judgment result of

[0081] Step S3-7: Generate a fake B image Input to the generator F of the second migration model B→A, and obtain the regenerated image regen of domain A A , the generated fake A image fake A Input to the generator F of the first migration model A→B , and obtain the regenerated image regen of domain B B , and calculate the cycle consistency loss:

[0082]

[0083] Among them, ||·||1 represents the L1 norm;

[0084] Step S3-8: Calculate the total loss based on the first adversarial loss, the second adversarial loss and the cycle consistency loss function, and correct the parameters of the migration model based on the total loss.

[0085] The total loss function is:

[0086] Loss(F A→B ,D B ,F B→A ,D A )=Loss GAN (F A→B ,D B ,A,B)+Loss GAN (F B→A ,D A ,B,A)+λLoss cyc (F A→B ,F B→A )

[0087] Among them: λ is an important parameter to balance the adversarial loss and cycle consistency loss.

[0088] Step S4: Use the test set to test the trained migration model, including:

[0089] Step S4-1: Input the real domain A image in the test sample into the generator F A→B Get the fake domain B image, and input the real domain B image in the test sample into the generator F B→A Get the fake domain A image,

[0090] Step S4-2: Calculate the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the fake domain A image and the true domain A image respectively, and calculate the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the true domain B and the fake domain B images respectively;

[0091] Among them, the mathematical expression of JS divergence is:

[0092]

[0093] Where: P n ,Q m They are fake A and real A The pictures in P n (i),Q m (i) is its normalized histogram, i.e., the probability distribution of each pixel value, To avoid division by zero, Q m (i) Smoothing or adding a small positive number ε to ensure that all values ​​are non-zero. Similarly, using JS divergence to measure fake B and real B The similarity D KL (fake B ||real B );

[0094] The mathematical expression of peak signal-to-noise ratio is:

[0095]

[0096] Where MAX is the maximum possible value of the pixel value (for example, for an 8-bit image, MAX=255; for a 16-bit image, MAX=65535);

[0097] The mathematical expression of the structural similarity index is:

[0098]

[0099] Where: μ x and μ y is a variable that measures brightness, and is a variable that measures contrast, σ xy is a variable to measure structural similarity, and C1 and C2 are constants to avoid the denominator being 0. When calculating SSIM, first calculate P n and Q m The SSIM value of a local window is then added and averaged to obtain P n and Q m The SSIM value of fake A and real A All images in the statistic are calculated and averaged to get the fake A and real A The overall SSIM value ssim(fake A ,real A ). Similarly, SSIM is used to measure fake B and real B The similarity of ssim(fake B,real B );

[0100] The mathematical expression of Frechette distance is:

[0101]

[0102] Among them, μ n and μ m Indicates P n and Q m The mean vector in the feature space extracted by the pre-trained Inception network, ∑ n and∑ m Indicates P n and Q m The covariance matrix in feature space, represents the square of the Euclidean distance between two mean vectors, and Tr(·) represents the trace of the matrix. Similarly, FID is used to measure fake B and real B The similarity of fid(fake B ,real B )

[0103] Step S4-3: Based on the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the fake domain A image and the true domain A image, combined with the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the true domain B and the fake domain B images, determine whether the migration model is trained.

[0104] Step S5: Based on the aggregate and gradation of the input image, the corresponding migration model is selected in combination with the aggregate and gradation of the image to be migrated, and the input image is input into the migration model to obtain the migrated asphalt mixture slice image output by the migration model. The result is as follows: Figure 4 shown.

[0105] Based on the above method, the comparison example is set as follows:

[0106] The overall structure is the same as that of the present embodiment, but the migration model adopts a series connection of 6 residual blocks, each of which consists of 2 3×3 convolutional layers, 2 normalization layers and 1 activation function layer.

[0107] The test results are as follows:

[0108] Table 1 Performance comparison between SIT-GAN migration model and original model (JS divergence)

[0109]

[0110] Table 2 Performance comparison between SIT-GAN migration model and original model (peak signal-to-noise ratio)

[0111]

[0112] Table 3 Performance comparison between SIT-GAN migration model and original model (structural similarity index)

[0113]

[0114] Table 4 Performance comparison between SIT-GAN migration model and original model (Fréchette distance)

[0115]

[0116]

[0117] In summary, this application aims to solve the pain points of high cost and difficulty in obtaining asphalt mixture slice images. Based on the image migration model, this application realizes the mutual migration between different types of slice images, effectively improving the data acquisition efficiency of mixture slice images. On the basis of the original CycleGAN model, this application introduces a self-designed multi-scale feature extraction module, adopts a multi-branch convolution splicing structure, and adds asymmetric convolution and deformable convolution. The above improvements enhance the migration model's ability to perceive and extract the complex multi-scale features of asphalt mixtures, and improve the quality of generated slice images.

[0118] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, RACd-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

Claims

1. A method for migrating asphalt mixture slice images based on an improved image translation neural network, characterized in that: include: Step S1: obtaining real slice images of asphalt mixtures with different gradations and aggregates; Step S2: cropping the acquired real slice image, and constructing an asphalt mixture dataset based on the cropped real slice image; Step S3: constructing a training set and a test set based on the asphalt mixture data set, and training a corresponding migration model based on the training set, wherein the migration model is constructed based on the SIT-GAN network; Step S4: Use the test set to test the trained migration model; Step S5: Based on the aggregate and gradation of the input image, the corresponding migration model is selected in combination with the aggregate and gradation of the image to be migrated, and the input image is input into the migration model to obtain the migrated asphalt mixture slice image output by the migration model.

2. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 1 is characterized in that: The step S1 comprises: Photos of all surfaces of a plurality of asphalt mixture specimens are obtained as real slice images, wherein the asphalt mixture specimens are polyhedrons, and at least one of the aggregate and the gradation of any two asphalt mixture specimens is different.

3. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 2 is characterized in that: The asphalt mixture specimen is a rectangular parallelepiped.

4. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 1 is characterized in that: The step S2 comprises: Step S2-1: removing invalid parts from all real slice images by clipping, and taking all clipped real slice images as original data sets, wherein the invalid parts at least include background areas, damaged areas, crack areas, reflective areas and stain areas, and the clipped real slice images are rectangular; Step S2-2: for all the real slice images in the original data set, each real slice image is cropped into at least one sub-slice image according to the aspect ratio of each real slice image, wherein the sub-slice image is a square, and the side length of any sub-slice image is consistent with the width of the corresponding real slice image; Step S2-3: All sub-slice images are combined with their aggregate and gradation information to form an element of the asphalt mixture data set, thereby obtaining the asphalt mixture data set.

5. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 4 is characterized in that: Among all the sub-slice images obtained by cropping the real slice images in the original data set, the similarity between any two sub-slice images is lower than a preconfigured first similarity threshold.

6. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 4 is characterized in that: The migration model includes a generator and a discriminator, and the generator includes: 1 reflection filling layer, used to extract and obtain the first feature map fm1 based on the sub-slice map; 1 convolution layer, 10 HFEM A modules, 1 activation function layer, 1 convolution layer, 10 HFEM B modules, 1 activation function layer and 1 convolution layer, 10 HFEM C modules, 1 activation function layer connected in sequence, used to extract and generate a second feature map fm2 based on the first feature map fm1; 3 layers of transposed convolution layers and activation function layers are used to further expand the size of the second feature map fm2 and compress the dimension to obtain the third feature map fm3; A reflection padding layer and a second 7×7 convolutional layer connected in sequence are used to obtain a migrated image based on the third feature map fm3.

7. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 6 is characterized in that: The step S3 comprises: Step S3-1: transform the size of all sub-slice images in the asphalt mixture dataset into input suitable for the migration model; Step S3-2: constructing a training set and a test set to construct a sample set based on the asphalt mixture data set, and dividing the sample set into a training set and a test set; Step S3-3: selecting a corresponding migration model according to the aggregate and gradation information of two sub-slice images of the training set, wherein at least one of the aggregate and gradation of the two sub-slice images is different; Step S3-4: Input the first sub-slice image to the generator F of the first migration model A→B , obtained by the generator F A→B Generated fake B-image B , and the generator F that feeds the second sub-slice graph into the second migration model B→A , obtained by the generator F B→A Generated fake A image fake A , where A is the domain of the first sub-slice map, representing the aggregate and gradation information of the first sub-slice map, and B is the domain of the second sub-slice map, representing the aggregate and gradation information of the second sub-slice map; Step S3-5: Combine the second sub-slice image and the obtained fake B image B Input to the discriminator D of the first migration model B , respectively get the real domain B image fake B The discrimination result of the fake domain B image and the discrimination result of the fake domain A image are obtained by combining the first sub-slice image with the fake domain A image fake A Input to the discriminator D of the second migration model A , respectively obtain the discrimination results of the real domain A image and the fake domain A image fake A The judgment result of Step S3-6: Based on the discrimination result of the real domain B image and the fake domain B image fake B The first adversarial loss is calculated based on the discrimination results of the real domain A image and the fake domain A image fake A The discrimination result calculates the second adversarial loss; Step S3-7: Generate fake B image fake B Input to the generator F of the second migration model B→A , and obtain the regenerated image regen of domain A A , the generated fake A image fake A Input to the generator F of the first migration model A→B , and obtain the regenerated image regen of domain B B , and calculate the cycle consistency loss; Step S3-8: Calculate the total loss based on the first adversarial loss, the second adversarial loss and the cycle consistency loss function, and correct the parameters of the migration model based on the total loss.

8. The asphalt mixture slice image migration method based on improved image translation neural network according to claim 7 is characterized in that: The step S4 comprises: Step S4-1: Input the real domain A image in the test sample into the generator F A→B Get the fake domain B image, and input the real domain B image in the test sample into the generator F B→A Get the fake domain A image, Step S4-2: Calculate the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the fake domain A image and the true domain A image respectively, and calculate the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the true domain B and the fake domain B images respectively; Step S4-3: Based on the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the fake domain A image and the true domain A image, combined with the JS divergence, peak signal-to-noise ratio, structural similarity index and Frechette distance of the true domain B and the fake domain B images, determine whether the migration model is trained.

9. An asphalt mixture slice image migration device based on an improved image translation neural network, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 8 is implemented.

Images