Pavement disease analysis method based on disease feature enhancement and image classification model

Abstract

The invention discloses a pavement disease analysis method based on disease feature enhancement and an image classification model, and belongs to the field of pavement disease analysis, and the method comprises the steps: obtaining a pavement image initial data set; the pavement image initial data set comprises a diseased pavement image and pavement disease information; constructing a pavement disease image classification model; training a pavement disease image classification model by using the pavement image initial data set to obtain a trained pavement disease image classification model; and obtaining a to-be-detected pavement disease image, and obtaining a disease classification result of the to-be-detected pavement disease image by using the trained pavement disease image classification model. According to the invention, the problem that the existing pavement disease identification method has high image requirements is solved.

Description

Technical Field

[0001] The present invention belongs to the field of pavement damage analysis, and in particular relates to a pavement damage analysis method based on damage feature enhancement and image classification model. Background Art

[0002] With the rapid development of my country's economy, the scale and complexity of the road traffic network are also increasing. The quality and safety of roads are of great significance to ensuring smooth traffic, reducing traffic accidents, and improving people's travel efficiency. However, due to the influence of various natural and human factors, various diseases will appear on the road surface, such as cracks, potholes, loose road surface, etc. These diseases will have a serious impact on the service life of the road and traffic safety.

[0003] Road diseases, such as cracks, potholes and loose pavement, not only affect the service life of the road, but also have a serious impact on road traffic safety. For example, if road cracks are not repaired in time, they may expand and eventually form potholes, which will cause serious obstacles to vehicle driving and may even cause traffic accidents. In addition, loose pavement may cause vehicles to lose control, thereby increasing the risk of traffic accidents.

[0004] Therefore, it is crucial to detect road damage in a timely and accurate manner and repair it. First, by repairing road damage in a timely manner, the damage can be prevented from further developing, thereby extending the service life of the road. For example, if cracks are repaired at an early stage, the cracks can be prevented from expanding, thereby avoiding larger-scale road damage. This will help increase the service life of the road, thereby reducing the construction cost of new roads.

[0005] Traditional road damage detection methods mainly rely on manual inspections, which are inefficient, costly, and difficult to meet the detection needs of large-scale road networks. In recent years, with the development of computer vision and deep learning technologies, the use of these technologies for automatic detection and identification of road damage has become a research hotspot. However, these methods usually require a large amount of labeled data, which is time-consuming and labor-intensive to obtain. In addition, due to the diversity and complexity of road images, how to improve the generalization ability of the model so that it can effectively identify road damage in a variety of different environments and conditions is also an important challenge. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a pavement disease analysis method based on disease feature enhancement and image classification model, which solves the problem that the existing pavement disease identification methods have high requirements on images.

[0007] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a pavement disease analysis method based on disease feature enhancement and image classification model, comprising:

[0008] Acquire an initial data set of road surface images; the initial data set of road surface images includes a damaged road surface image and road surface damage information;

[0009] Constructing a pavement disease image classification model; using an initial pavement image data set to train the pavement disease image classification model to obtain a trained pavement disease image classification model; the pavement disease image classification model includes a disease feature enhancement module, a feature map cleaning module, a feature map clustering module, a feature extraction module, a plurality of texture enhancement modules and a classifier;

[0010] The disease feature enhancement module is used to enhance the disease features of the diseased road surface images in the initial road surface image data set to obtain an enhanced set of diseased road surface images;

[0011] The feature image cleaning module is used to remove interference features according to the enhanced diseased road surface image set to obtain a cleaned diseased road surface image set;

[0012] The feature graph clustering module is used to divide the cleaned diseased pavement image set into an easy-to-separate image subset and a hard-to-separate image subset; the hard-to-separate image subset includes a number of hard-to-separate image clusters;

[0013] The feature extraction module is used to extract features from each image in the easily separable image subset to obtain initial features of each image in the easily separable image subset;

[0014] Each of the texture enhancement modules is used to perform texture enhancement on the corresponding hard-to-distinguish image clusters to obtain final texture features of each image in the hard-to-distinguish image clusters;

[0015] The classifier is used to classify the defects according to the initial features of each image in the easy-to-separate image subset and the final texture features of each image in each hard-to-separate image cluster;

[0016] The pavement disease image to be tested is obtained, and the disease classification result of the pavement disease image to be tested is obtained by using the trained pavement disease image classification model.

[0017] Furthermore, the defect features of the defective road surface images in the initial road surface image data set are enhanced to obtain an enhanced set of defective road surface images, specifically:

[0018] Acquire the current diseased road surface image from the initial road surface image data set;

[0019] Performing grayscale processing on the current diseased road surface image to obtain a grayscale image of the current diseased road surface image;

[0020] According to the grayscale image of the current diseased road surface image, the current diseased road surface image is divided into image blocks to obtain an image block division result of the current diseased road surface image;

[0021] According to the image block division result of the current diseased road surface image, feature enhancement is performed to obtain an enhanced diseased road surface image of the current diseased road surface image;

[0022] The next damaged road surface image is subjected to disease feature enhancement until enhanced damaged road surface images of all damaged road surface images in the initial road surface image data set are obtained, thereby obtaining an enhanced damaged road surface image set.

[0023] Furthermore, the image blocks of the current diseased road surface image are divided according to the grayscale image of the current diseased road surface image to obtain the image block division result of the current diseased road surface image, which is specifically:

[0024] Use the target detection algorithm to identify the target of the current damaged road surface image and obtain the damaged area frame;

[0025] Set the contour magnification parameters; based on the contour magnification parameters, expand the disease area frame outward to obtain an enlarged disease area frame;

[0026] Cut the enlarged damaged area frame from the grayscale image of the current damaged road surface image to obtain a background area without damage;

[0027] According to the enlarged diseased area frame and the disease-free background area, the image block division result of the current diseased road surface image is obtained.

[0028] Furthermore, the feature enhancement is performed according to the image block division result of the current damaged road surface image to obtain an enhanced damaged road surface image of the current damaged road surface image, specifically:

[0029] The grayscale value of each pixel in the disease-free background area is reduced to obtain a processed disease-free background area; the grayscale value of each pixel in the disease-free background area after the processing is:

[0030]

[0031] Where F(i) is the gray value of pixel i in the disease-free background area after treatment; H i is the gray value of pixel i in the disease-free background area; Z(H i ) is the gray value of the background area without disease H i R is the total number of pixels in the disease-free background area;

[0032] The processed disease-free background area and the enlarged disease area frame are superimposed on one image to obtain an integrated diseased road surface image;

[0033] The integrated damaged pavement image is converted into an RGB image to obtain an enhanced damaged pavement image of the current damaged pavement image.

[0034] Furthermore, the expression of the cleaned damaged road surface image in the cleaned damaged road surface image set is:

[0035]

[0036] Among them, X” is the cleaned damaged road surface image; f() is the binarization operation; g() is the debinarization operation; X' is the enhanced damaged road surface image after the closing operation; S2 and S1 are both structural elements; X is the enhanced damaged road surface image; ⊕ is the expansion operation; For corrosion operation.

[0037] Furthermore, the cleaned diseased road surface image set is divided into an easily separable image subset and a difficult-to-separate image subset, specifically: obtaining the initial features of each cleaned diseased road surface image in the cleaned diseased road surface image set, and considering the initial features of each cleaned diseased road surface image as clusters; obtaining discrete clusters, and dividing the cleaned diseased road surface images belonging to the discrete clusters into the easily separable image subset; eliminating the clusters of each image in the easily separable image subset to obtain the remaining clusters; and calculating the distances between the clusters in the remaining clusters; merging the clusters that meet the merging conditions until the number of clusters is less than or equal to the preset number of clusters, to obtain a number of difficult-to-separate image clusters; the expression of the distance between the clusters is:

[0038]

[0039] Where D is the distance between clusters; min is the minimum function; max is the maximum function; d is the distance calculation function; V A is the initial feature vector of the image in cluster A; V B is the initial feature vector of the image in cluster B; is the average value of the distance between the initial feature vectors of the images in cluster A and cluster B.

[0040] Further, the texture enhancement module includes a texture feature extractor, a semantic feature extractor and a weighted fusion unit;

[0041] The texture feature extractor is used to extract local texture features of each image in the hard-to-distinguish image cluster;

[0042] The semantic feature extractor is used to extract the semantic features of each image in the hard-to-distinguish image cluster;

[0043] The weighted fusion unit is used to perform weighted fusion on the texture features and semantic features of each image in the hard-to-distinguish image cluster to obtain the final texture features of each image in the hard-to-distinguish image cluster.

[0044] Furthermore, the expression of the final texture feature is:

[0045] F T '=Resnet18([F LBP ,F Con ,F Cor ,F ASM ,F Dis ,F Hom ,F Ent ])

[0046] F S =S(F GLCM ),S∈{Con,Cor,ASM,Dis,Hom,Ent}

[0047] Among them, F T ' is the final texture feature; Resnet18 is the pre-trained ResNet18 network; F LBP is the local texture feature; F Con is the contrast feature; F Cor is the correlation feature; F ASM is the angular second-order moment characteristic; F Dis is a differential feature; F Hom F is a homogeneous characteristic; Ent is the entropy characteristic; F S is the total feature map; S is the feature selection operation; F GLCM is a semantic feature; Con is a contrast marker; Cor is a correlation marker; ASM is an angular second moment marker; Dis is a difference marker; Hom is a homogeneity marker; Ent is an entropy marker.

[0048] Furthermore, the classifier includes a common classification head for classifying defects according to initial features of each image in the easy-to-classify image subset and N special classification heads for classifying defects according to final texture features of each image in each hard-to-classify image cluster.

[0049] Furthermore, the pavement disease image to be tested is obtained, and the disease classification result of the pavement disease image to be tested is obtained by using the trained pavement disease image classification model, which is specifically:

[0050] Obtain the pavement disease image to be tested, and use the disease feature enhancement module and the feature map cleaning module to obtain the cleaned pavement disease image to be tested; calculate the distance between the cleaned pavement disease image to be tested and the easy-to-separate image subset and each difficult-to-separate image cluster divided in the feature map clustering module, and select the one with the smallest distance as the subsequent input module; if the distance with the easy-to-separate image subset is the smallest, input the cleaned pavement disease image to be tested into the feature extraction module, and use the ordinary classification head to obtain the disease classification result of the pavement disease image to be tested; otherwise, input the cleaned pavement disease image to be tested into the texture enhancement module with the smallest distance, and use the corresponding special classification head to obtain the disease classification result of the pavement disease image to be tested.

[0051] The beneficial effects of the present invention are as follows: the method reduces the image background information in the training set, thereby achieving disease feature enhancement, so that the trained pavement disease image classification model can read more obvious disease features, thereby achieving classification; the same operation is performed on the detection, first obtaining the target frame, and then enhancing the contrast with the background, thereby achieving disease feature enhancement. Such processing avoids the image classification model's dependence on image quality; at the same time, during the training process, the training set is firstly roughly classified, and at the same time, different classification channels are divided in the pavement disease image classification model, so that each channel achieves the best classification accuracy in the corresponding image cluster, thereby avoiding errors caused by large data volume training. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The figure is a flow chart of the method of the present invention.

[0053] Figure 2 It is a model framework diagram of the present invention. DETAILED DESCRIPTION

[0054] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0055] like Figure 1 and Figure 2 As shown, in one embodiment of the present invention, a pavement disease analysis method based on disease feature enhancement and image classification model includes:

[0056] Acquire an initial data set of road surface images; the initial data set of road surface images includes a damaged road surface image and road surface damage information;

[0057] Constructing a pavement disease image classification model; using an initial pavement image data set to train the pavement disease image classification model to obtain a trained pavement disease image classification model; the pavement disease image classification model includes a disease feature enhancement module, a feature map cleaning module, a feature map clustering module, a feature extraction module, a plurality of texture enhancement modules and a classifier;

[0058] The disease feature enhancement module is used to enhance the disease features of the diseased road surface images in the initial road surface image data set to obtain an enhanced set of diseased road surface images;

[0059] The feature image cleaning module is used to remove interference features according to the enhanced diseased road surface image set to obtain a cleaned diseased road surface image set;

[0060] The feature graph clustering module is used to divide the cleaned diseased pavement image set into an easy-to-separate image subset and a hard-to-separate image subset; the hard-to-separate image subset includes a number of hard-to-separate image clusters;

[0061] The feature extraction module is used to extract features from each image in the easily separable image subset to obtain initial features of each image in the easily separable image subset;

[0062] Each of the texture enhancement modules is used to perform texture enhancement on the corresponding hard-to-distinguish image clusters to obtain final texture features of each image in the hard-to-distinguish image clusters;

[0063] The classifier is used to classify the defects according to the initial features of each image in the easy-to-separate image subset and the final texture features of each image in each hard-to-separate image cluster;

[0064] The pavement disease image to be tested is obtained, and the disease classification result of the pavement disease image to be tested is obtained by using the trained pavement disease image classification model.

[0065] In this embodiment, the feature extraction module recognizes the pavement disease image based on the Transformer model combined with traditional convolution, and uses the traditional convolution layer to extract features from the preprocessed pavement image. The convolution layer can capture the local features of the image, including texture, shape, etc. By adjusting the size and number of convolution kernels, the extraction of features of different scales can be achieved. The output of the feature fusion layer is analyzed using the self-attention mechanism using the Transformer layer. Through the self-attention mechanism, the model can automatically learn important areas in the image and give them greater attention. This layer can capture the global features of the image. The Transformer can divide the pavement image into multiple small blocks for feature extraction and conversion, and use Patch Embedding to embed each small block into a vector; at the same time, the depth and width of the Transformer are retained, so that it can better capture the features and patterns of the image.

[0066] In convolutional neural networks (CNNs), convolution operations are good at extracting local features, but they are still limited in capturing global feature representations. In Vision Transformer, the cascaded self-attention module can capture long-range feature dependencies, but the long-range attention mechanism ignores the details of local features. This project is based on a multi-layer Transformer, inserting traditional convolutions between layers of the network for fine-grained feature extraction. CNN is used to extract local features of the image at the low level, while Transformer is used to capture global features at the high level. By fusing features of different scales, the aim is to hierarchically extract local and global features of diseased pavement images through convolution + global attention, thereby improving model performance and computational overhead.

[0067] The defect features of the defective road surface images in the initial road surface image data set are enhanced to obtain an enhanced defective road surface image set, specifically:

[0068] Acquire the current diseased road surface image from the initial road surface image data set;

[0069] Performing grayscale processing on the current diseased road surface image to obtain a grayscale image of the current diseased road surface image;

[0070] According to the grayscale image of the current diseased road surface image, the current diseased road surface image is divided into image blocks to obtain an image block division result of the current diseased road surface image;

[0071] According to the image block division result of the current diseased road surface image, feature enhancement is performed to obtain an enhanced diseased road surface image of the current diseased road surface image;

[0072] The next damaged road surface image is subjected to disease feature enhancement until enhanced damaged road surface images of all damaged road surface images in the initial road surface image data set are obtained, thereby obtaining an enhanced damaged road surface image set.

[0073] The image blocks of the current diseased road surface image are divided according to the grayscale image of the current diseased road surface image to obtain the image block division result of the current diseased road surface image, which is specifically:

[0074] Use the target detection algorithm to identify the target of the current damaged road surface image and obtain the damaged area frame;

[0075] Set the contour magnification parameters; based on the contour magnification parameters, expand the disease area frame outward to obtain an enlarged disease area frame;

[0076] Cut the enlarged damaged area frame from the grayscale image of the current damaged road surface image to obtain a background area without damage;

[0077] According to the enlarged diseased area frame and the disease-free background area, the image block division result of the current diseased road surface image is obtained.

[0078] The feature enhancement is performed according to the image block division result of the current diseased road surface image to obtain the enhanced diseased road surface image of the current diseased road surface image, specifically:

[0079] The grayscale value of each pixel in the disease-free background area is reduced to obtain a processed disease-free background area; the grayscale value of each pixel in the disease-free background area after the processing is:

[0080]

[0081] Where F(i) is the gray value of pixel i in the disease-free background area after treatment; H i is the gray value of pixel i in the disease-free background area; Z(H i ) is the gray value of the background area without disease H i R is the total number of pixels in the disease-free background area;

[0082] The processed disease-free background area and the enlarged disease area frame are superimposed on one image to obtain an integrated diseased road surface image;

[0083] The integrated damaged pavement image is converted into an RGB image to obtain an enhanced damaged pavement image of the current damaged pavement image.

[0084] In this embodiment, the set contour magnification parameter is used to expand the original defect area frame outward as a whole, which can be used to avoid cutting the defect edge during subsequent grayscale weakening, thereby causing errors in pavement defect feature extraction.

[0085] In this embodiment, the grayscale of the background area is reduced proportionally as a whole, so that the characteristics of the road surface damage can be highlighted while retaining certain background features.

[0086] In this embodiment, grayscale reduction is performed according to the numerical ratio of grayscale values, which can retain the global information of the background to the greatest extent and highlight the characteristics of road surface damage.

[0087] The expression of the cleaned damaged road surface image in the cleaned damaged road surface image set is:

[0088]

[0089] Among them, X” is the cleaned damaged road surface image; f() is the binarization operation; g() is the debinarization operation; X' is the enhanced damaged road surface image after the closing operation; S2 and S1 are both structural elements; X is the enhanced damaged road surface image; ⊕ is the expansion operation; For corrosion operation.

[0090] The cleaned diseased road surface image set is divided into an easy-to-separate image subset and a difficult-to-separate image subset, specifically: obtaining the initial features of each cleaned diseased road surface image in the cleaned diseased road surface image set, and considering the initial features of each cleaned diseased road surface image as clusters; obtaining discrete clusters, and dividing the cleaned diseased road surface images belonging to the discrete clusters into the easy-to-separate image subset; eliminating the clusters of each image in the easy-to-separate image subset to obtain the remaining clusters; and calculating the distances between the clusters in the remaining clusters; merging the clusters that meet the merging conditions until the number of clusters is less than or equal to the preset number of clusters, to obtain a number of difficult-to-separate image clusters; the expression of the distance between the clusters is:

[0091]

[0092] Where D is the distance between clusters; min is the minimum function; max is the maximum function; d is the distance calculation function; V A is the initial feature vector of the image in cluster A; V B is the initial feature vector of the image in cluster B; is the average value of the distance between the initial feature vectors of the images in cluster A and cluster B.

[0093] The texture enhancement module includes a texture feature extractor, a semantic feature extractor and a weighted fusion unit;

[0094] The texture feature extractor is used to extract local texture features of each image in the hard-to-distinguish image cluster;

[0095] The semantic feature extractor is used to extract the semantic features of each image in the hard-to-distinguish image cluster;

[0096] The weighted fusion unit is used to perform weighted fusion on the texture features and semantic features of each image in the hard-to-distinguish image cluster to obtain the final texture features of each image in the hard-to-distinguish image cluster.

[0097] The expression of the final texture feature is:

[0098] F T '=Resnet18([F LBP ,F Con ,F Cor ,F ASM ,F Dis ,F Hom ,F Ent ])

[0099] F S =S(F GLCM ),S∈{Con,Cor,ASM,Dis,Hom,Ent}

[0100] Among them, F T ' is the final texture feature; Resnet18 is the pre-trained ResNet18 network; F LBP is the local texture feature; F Con is the contrast feature; F Cor is the correlation feature; F ASM is the angular second-order moment characteristic; F Dis is a differential feature; F Hom F is a homogeneous characteristic; Ent is the entropy characteristic; F S is the total feature map; S is the feature selection operation; F GLCM is a semantic feature; Con is a contrast marker; Cor is a correlation marker; ASM is an angular second moment marker; Dis is a difference marker; Hom is a homogeneity marker; Ent is an entropy marker.

[0101] The classifier includes a common classification head for classifying defects according to initial features of each image in the easy-to-classify image subset and N special classification heads for classifying defects according to final texture features of each image in each hard-to-classify image cluster.

[0102] The pavement disease image to be tested is obtained, and the disease classification result of the pavement disease image to be tested is obtained by using the trained pavement disease image classification model, specifically:

[0103] Obtain the pavement disease image to be tested, and use the disease feature enhancement module and the feature map cleaning module to obtain the cleaned pavement disease image to be tested; calculate the distance between the cleaned pavement disease image to be tested and the easy-to-separate image subset and each difficult-to-separate image cluster divided in the feature map clustering module, and select the one with the smallest distance as the subsequent input module; if the distance with the easy-to-separate image subset is the smallest, input the cleaned pavement disease image to be tested into the feature extraction module, and use the ordinary classification head to obtain the disease classification result of the pavement disease image to be tested; otherwise, input the cleaned pavement disease image to be tested into the texture enhancement module with the smallest distance, and use the corresponding special classification head to obtain the disease classification result of the pavement disease image to be tested.

Claims

1. A pavement disease analysis method based on disease feature enhancement and image classification model, characterized in that: include: Acquire an initial data set of road surface images; the initial data set of road surface images includes a damaged road surface image and road surface damage information; Construct a pavement damage image classification model; Using the initial pavement image data set to train a pavement disease image classification model, a trained pavement disease image classification model is obtained; the pavement disease image classification model includes a disease feature enhancement module, a feature map cleaning module, a feature map clustering module, a feature extraction module, a plurality of texture enhancement modules and a classifier; The disease feature enhancement module is used to enhance the disease features of the diseased road surface images in the initial road surface image data set to obtain an enhanced set of diseased road surface images; The feature image cleaning module is used to remove interference features according to the enhanced diseased road surface image set to obtain a cleaned diseased road surface image set; The feature graph clustering module is used to divide the cleaned diseased pavement image set into an easy-to-separate image subset and a hard-to-separate image subset; the hard-to-separate image subset includes a number of hard-to-separate image clusters; The feature extraction module is used to extract features from each image in the easily separable image subset to obtain initial features of each image in the easily separable image subset; Each of the texture enhancement modules is used to perform texture enhancement on the corresponding hard-to-distinguish image clusters to obtain final texture features of each image in the hard-to-distinguish image clusters; The classifier is used to classify the defects according to the initial features of each image in the easy-to-separate image subset and the final texture features of each image in each hard-to-separate image cluster; The pavement disease image to be tested is obtained, and the disease classification result of the pavement disease image to be tested is obtained by using the trained pavement disease image classification model.

2. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 1 is characterized in that: The defect features of the defective road surface images in the initial road surface image data set are enhanced to obtain an enhanced defective road surface image set, specifically: Acquire the current diseased road surface image from the initial road surface image data set; Performing grayscale processing on the current diseased road surface image to obtain a grayscale image of the current diseased road surface image; According to the grayscale image of the current diseased road surface image, the current diseased road surface image is divided into image blocks to obtain an image block division result of the current diseased road surface image; According to the image block division result of the current diseased road surface image, feature enhancement is performed to obtain an enhanced diseased road surface image of the current diseased road surface image; The next damaged road surface image is subjected to disease feature enhancement until enhanced damaged road surface images of all damaged road surface images in the initial road surface image data set are obtained, thereby obtaining an enhanced damaged road surface image set.

3. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 2 is characterized in that: The image blocks of the current diseased road surface image are divided according to the grayscale image of the current diseased road surface image to obtain the image block division result of the current diseased road surface image, which is specifically: Use the target detection algorithm to identify the target of the current damaged road surface image and obtain the damaged area frame; Set the contour magnification parameters; based on the contour magnification parameters, expand the disease area frame outward to obtain an enlarged disease area frame; Cut the enlarged damaged area frame from the grayscale image of the current damaged road surface image to obtain a background area without damage; According to the enlarged diseased area frame and the disease-free background area, the image block division result of the current diseased road surface image is obtained.

4. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 3 is characterized in that: The feature enhancement is performed according to the image block division result of the current diseased road surface image to obtain the enhanced diseased road surface image of the current diseased road surface image, specifically: The grayscale value of each pixel in the disease-free background area is reduced to obtain a processed disease-free background area; the grayscale value of each pixel in the disease-free background area after the processing is: Where F(i) is the gray value of pixel i in the disease-free background area after treatment; H i is the gray value of pixel i in the disease-free background area; Z(H i ) is the gray value of the background area without disease H i R is the total number of pixels in the disease-free background area; The processed disease-free background area and the enlarged disease area frame are superimposed on one image to obtain an integrated diseased road surface image; The integrated damaged pavement image is converted into an RGB image to obtain an enhanced damaged pavement image of the current damaged pavement image.

5. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 1 is characterized in that: The expression of the cleaned damaged road surface image in the cleaned damaged road surface image set is: Among them, X” is the cleaned damaged road surface image; f() is the binarization operation; g() is the debinarization operation; X’ is the enhanced damaged road surface image after the closing operation; S2 and S1 are both structural elements; X is the enhanced damaged road surface image; For expansion operation; For corrosion operation.

6. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 1 is characterized in that: The method of dividing the cleaned diseased pavement image set into an easily separable image subset and a difficult-to-separate image subset is specifically as follows: obtaining initial features of each cleaned diseased pavement image in the cleaned diseased pavement image set, and treating the initial features of each cleaned diseased pavement image as a cluster; obtaining discrete clusters, and dividing the cleaned diseased pavement images belonging to the discrete clusters into the easily separable image subset; and removing the clusters of each image in the easily separable image subset to obtain the remaining clusters; And calculate the distance between each cluster in the remaining clusters; merge the clusters that meet the merging conditions until the number of clusters is less than or equal to the preset number of clusters, and obtain several difficult-to-distinguish image clusters; the expression of the distance between the clusters is: Where D is the distance between clusters; min is the minimum function; max is the maximum function; d is the distance calculation function; V A is the initial feature vector of the image in cluster A; V B is the initial feature vector of the image in cluster B; is the average value of the distance between the initial feature vectors of the images in cluster A and cluster B.

7. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 1 is characterized in that: The texture enhancement module includes a texture feature extractor, a semantic feature extractor and a weighted fusion unit; The texture feature extractor is used to extract local texture features of each image in the hard-to-distinguish image cluster; The semantic feature extractor is used to extract the semantic features of each image in the hard-to-distinguish image cluster; The weighted fusion unit is used to perform weighted fusion on the texture features and semantic features of each image in the hard-to-distinguish image cluster to obtain the final texture features of each image in the hard-to-distinguish image cluster.

8. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 7 is characterized in that: The expression of the final texture feature is: F T '=Resnet18([F LBP ,F Con ,F Cor ,F ASM ,F Dis ,F Hom ,F Ent ]) F S =S(F GLCM ),S∈{Con,Cor,ASM,Dis,Hom,Ent} Among them, F T ' is the final texture feature; Resnet18 is the pre-trained ResNet18 network; F LBP is the local texture feature; F Con is the contrast feature; F Cor is the correlation feature; F ASM is the angular second-order moment characteristic; F Dis is a differential feature; F Hom F is a homogeneous characteristic; Ent is the entropy characteristic; F S is the total feature map; S is the feature selection operation; F GLCM is a semantic feature; Con is a contrast marker; Cor is a correlation marker; ASM is an angular second moment marker; Dis is a difference marker; Hom is a homogeneity marker; Ent is an entropy marker.

9. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 1 is characterized in that: The classifier includes a common classification head for classifying defects according to initial features of each image in the easy-to-classify image subset and N special classification heads for classifying defects according to final texture features of each image in each hard-to-classify image cluster.

10. The pavement disease analysis method based on disease feature enhancement and image classification model according to claim 1 is characterized in that: The pavement disease image to be tested is obtained, and the disease classification result of the pavement disease image to be tested is obtained by using the trained pavement disease image classification model, specifically: Obtain the road surface disease image to be tested, and use the disease feature enhancement module and the feature map cleaning module to obtain the cleaned road surface disease image to be tested; calculate the distance between the cleaned road surface disease image to be tested and the easy-to-separate image subsets and each difficult-to-separate image clusters divided by the feature map clustering module, and select the one with the smallest distance as the subsequent input module; If the distance to the easily distinguishable image subset is the smallest, the cleaned pavement disease image to be tested is input into the feature extraction module, and the ordinary classification head is used to obtain the disease classification result of the pavement disease image to be tested; otherwise, the cleaned pavement disease image to be tested is input into the texture enhancement module with the smallest distance, and the corresponding special classification head is used to obtain the disease classification result of the pavement disease image to be tested.

Images