A high-speed pavement crack detection method based on multi-dimensional feature extraction

Abstract

The present application relates to a high-speed pavement crack detection method with multi-dimensional feature extraction, which uses a filtering method to pre-process the pavement image, filter the noise in the image and retain the texture details and boundary information of the crack; uses a fractal feature method to construct a crack feature map to capture the shape and texture features of the crack, and uses an MSF feature extraction network to extract crack features at different scales; constructs a feature fusion network to fuse feature information at different scales, and adds IDS convolution to enhance features and reduce the amount of parameters and calculations; fuses the feature maps of each layer through a cluster network to form a crack detection map, thereby achieving efficient and accurate pavement crack detection. The present application has a high crack detection accuracy, strong robustness and good model generalization ability, and has better practical significance.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a high-speed pavement crack detection method based on multi-dimensional feature extraction. Background Art

[0002] Intelligent transportation systems are becoming increasingly important in areas such as road safety enhancement, traffic flow control, and infrastructure maintenance. Among them, information from crack detection is crucial for evaluating current road conditions and predicting road service life, and is a key link in intelligent transportation systems. However, current automated crack detection methods still face multiple challenges. The complex and changeable road image background, including lighting fluctuations, shadow interference, and uneven road texture and color, greatly increase the difficulty of crack feature extraction. In addition, cracks have different morphologies, and their sizes, shapes, and directions vary greatly, requiring detection methods to have good adaptability and robustness. Existing methods have made some progress in feature characterization, but there are still limitations in processing multi-scale crack features and contextual information. Summary of the invention

[0003] In order to overcome at least one deficiency in the prior art, the present application provides a high-speed pavement crack detection method based on multi-dimensional feature extraction.

[0004] In a first aspect, a high-speed pavement crack detection method based on multi-dimensional feature extraction is provided, comprising:

[0005] The acquired highway road image is subjected to denoising to obtain a denoised road image; the denoising process uses a filtering method to filter out medium and high frequency information in the highway road image;

[0006] According to the final pixel value of each pixel in the denoised road surface image, the fractal feature and the fractal dimension eigenvalue of each pixel are determined, and a single feature representation of each pixel is obtained according to the fractal feature and the fractal dimension eigenvalue; the single feature representations of all pixels constitute a crack feature map;

[0007] The crack feature map is input into the MSF feature extraction network to obtain the fractal features of cracks at different scales;

[0008] The fractal features of cracks at different scales are input into the feature fusion network for feature fusion to obtain fusion features at different scales.

[0009] The fused features of different scales are input into the cluster network to obtain the crack detection map of the highway pavement image; the cluster network groups the fused features of different scales according to the scale size, and uses different convolution kernels to perform convolution operations on the fused features in different groups to obtain the convolution results; the convolution results corresponding to different groups are added element by element to obtain the crack detection map of the highway pavement image; the grouping according to the scale size includes: sorting the fused features of different scales from small to large according to the scale size, grouping the first three fused features after sorting into one group, and grouping the remaining fused features into one group.

[0010] In one embodiment, performing denoising on the acquired highway road image to obtain a denoised road image includes:

[0011] Calculate the filter weight corresponding to each pixel in the highway road image;

[0012] According to the filtering weights, the pixel values ​​of the pixels in the neighborhood in four directions around any pixel are weighted and summed to obtain the weighted result of the pixel in each direction;

[0013] The weighted results in all directions are averaged as the final pixel value of the pixel;

[0014] The final pixel values ​​of all pixels constitute the denoised road image.

[0015] In one embodiment, the filter weight corresponding to each pixel in the highway road image is calculated using the following formula:

[0016]

[0017] in, Pixel The corresponding filter weights are represents the Gamma function, is the index of the pixel, is the order, Represents factorial.

[0018] In one embodiment, according to the filtering weights, the pixel values ​​of the pixels in the neighborhood in four directions around any pixel are weighted and summed to obtain the weighted result of the pixel in each direction, using the following formula:

[0019]

[0020]

[0021]

[0022]

[0023] in, , , , Represents pixel points ( x,y ) is the weighted result above, below, left, and right, x,y The pixels are x Axis coordinates and y Axis coordinates; is the index of the pixel, is the number of pixels in the neighborhood, Pixel The corresponding filter weights are Represents a pixel ( ), Represents pixel The pixel value of Represents a pixel ( ), Represents pixel The pixel value of .

[0024] In one embodiment, the fractal feature and fractal dimension feature value of each pixel are determined according to the final pixel value of each pixel in the denoised road surface image, including:

[0025] Determine the fractal dimension eigenvalue of each pixel using the following formula:

[0026]

[0027] in, Pixel The fractal dimension eigenvalue of is a pixel ( x,y ), Pixel The average value of the final pixel value of the pixels in the neighborhood with a radius of r. is the normalized coefficient index; Indicates absolute value;

[0028] To determine the fractal characteristics of each pixel, the following formula is used:

[0029]

[0030] in, Pixel The fractal characteristics of x,y The pixels are x Axis coordinates and y Axis coordinates, Pixel The fractal dimension eigenvalue of Pixel The set of pixels in a neighborhood with a radius of r. express Norm.

[0031] In one embodiment, the MSF feature extraction network includes five feature extraction layers of different scales. The crack fractal features output by the feature extraction layer with a higher scale in two adjacent feature extraction layers are input into the feature extraction layer with a lower scale after downsampling; the five feature extraction layers of different scales respectively output a crack fractal feature.

[0032] In one embodiment, the feature fusion network includes a convolution module, a feature enhancement module and a fusion module;

[0033] The convolution module is used to perform convolution operations on crack fractal features of different scales to obtain multiple convolution results;

[0034] The feature enhancement module is used to enhance the context information of the lowest-scale crack fractal features to obtain enhanced features;

[0035] The fusion module is used to fuse the enhanced features with multiple convolution results to obtain fused features of different scales.

[0036] In one embodiment, the crack fractal features of different scales include 5, which are denoted as L1, L2, L3, L4, and L5, wherein the scales of L1, L2, L3, L4, and L5 increase in sequence;

[0037] Enhance the feature of context information of L1 to obtain the enhanced feature C;

[0038] Convolution operations are performed on L1, L2, L3, L4, and L5 respectively to obtain multiple convolution results, which are recorded as M1, M2, M3, M4, and M5;

[0039] C and M1 are added element by element to obtain the fused feature P1;

[0040] P1 and M2 are added element by element to obtain the fused feature P2;

[0041] P2 and M3 are added element by element to obtain the fused feature P3;

[0042] P3 and M4 are added element by element to obtain the fused feature P4;

[0043] P4 and M5 are added element by element to obtain the fused feature P5.

[0044] In one embodiment, the feature enhancement module is used to:

[0045] Perform convolution operation on the lowest-scale crack fractal features to extract local context information;

[0046] Perform global maximum pooling on the lowest-scale crack fractal features to extract global context information;

[0047] Add the local context information and the global context information element by element to obtain the context information feature map P;

[0048] The context information feature map P is convolved with a convolution rate of 1 and a convolution rate of 3 to obtain two crack feature maps f1 and f2;

[0049] Add the crack feature map f1 and the crack feature map f2 element by element to obtain a new crack feature map F;

[0050] Perform global maximum pooling, convolution, and Sigmoid activation function processing on the new crack feature map F to obtain channel weights;

[0051] The crack feature map f1 and the crack feature map f2 are respectively multiplied element by element with the channel weight to obtain the feature map f1' and the feature map f2' respectively;

[0052] The feature map f1' and the feature map f2' are concatenated to obtain the enhanced feature F'.

[0053] In one embodiment, the method further comprises:

[0054] IDS convolution is performed on the fusion features of different scales respectively to obtain the fusion features of different scales after feature enhancement.

[0055] In a second aspect, a high-speed pavement crack detection device with multi-dimensional feature extraction is provided, comprising:

[0056] A denoising module is used to perform denoising on the acquired highway road image to obtain a denoised road image; the denoising process uses a filtering method to filter out medium and high frequency information in the highway road image;

[0057] A crack feature map acquisition module is used to determine the fractal feature and fractal dimension feature value of each pixel point according to the final pixel value of each pixel point in the denoised road surface image, and obtain a single feature representation of each pixel point according to the fractal feature and fractal dimension feature value; the single feature representations of all pixels points constitute a crack feature map;

[0058] The feature extraction module is used to input the crack feature map into the MSF feature extraction network to obtain the fractal features of cracks at different scales;

[0059] The feature fusion module is used to input the fractal features of cracks of different scales into the feature fusion network for feature fusion to obtain fusion features of different scales;

[0060] The detection module is used to input fusion features of different scales into the cluster network to obtain a crack detection map of the highway pavement image; the cluster network groups the fusion features of different scales according to the scale size, and uses different convolution kernels to perform convolution operations on the fusion features in different groups to obtain convolution results; the convolution results corresponding to different groups are added element by element to obtain a crack detection map of the highway pavement image; grouping according to scale size includes: sorting the fusion features of different scales from small to large according to the scale size, grouping the first three fusion features after sorting into one group, and grouping the remaining fusion features into one group.

[0061] Compared with the prior art, the present application has the following beneficial effects: the high-speed pavement crack detection method of the present application with multi-dimensional feature extraction uses a filtering method to pre-process the pavement image, filters the noise in the image and retains the texture details and boundary information of the crack; uses a fractal feature method to construct a crack feature map to capture the shape and texture features of the crack, and uses an MSF feature extraction network to extract crack features at different scales; constructs a feature fusion network to fuse feature information at different scales, and adds IDS convolution to enhance features and reduce the amount of parameters and calculations; fuses the feature maps of each layer through a cluster network to form a crack detection map, thereby achieving efficient and accurate pavement crack detection. The present application has a high crack detection accuracy, strong robustness and good model generalization ability, and has better practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present application may be better understood by referring to the following description given in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification. In the drawings:

[0063] Figure 1 A flowchart of a high-speed pavement crack detection method based on multi-dimensional feature extraction is shown;

[0064] Figure 2 Shows the structural schematic diagram of the MSF feature extraction network;

[0065] Figure 3 The structural diagram of the feature fusion network is shown;

[0066] Figure 4 shows a schematic structural diagram of a feature enhancement module;

[0067] Figure 5 shows a schematic diagram of the structure of a cluster network;

[0068] Figure 6 The structural block diagram of the high-speed pavement crack detection device with multi-dimensional feature extraction is shown. DETAILED DESCRIPTION

[0069] The exemplary embodiments of the present application will be described below in conjunction with the accompanying drawings. For the sake of clarity and conciseness, not all features of the actual embodiments are described in the specification. However, it should be understood that many implementation-specific decisions can be made in the process of developing any such actual embodiments in order to achieve the specific goals of the developer, and these decisions may vary from embodiment to embodiment.

[0070] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, only the device structure closely related to the scheme according to the present application is shown in the drawings, while other details that are not closely related to the present application are omitted.

[0071] It should be understood that the present application is not limited to the described implementation forms due to the following description with reference to the accompanying drawings. In this article, where feasible, the embodiments can be combined with each other, features between different embodiments can be replaced or borrowed, and one or more features can be omitted in one embodiment.

[0072] The present application provides a method for detecting cracks in a high-speed road surface by extracting multi-dimensional features. Figure 1 A flowchart of a high-speed pavement crack detection method based on multi-dimensional feature extraction is shown in FIG. Figure 1 , methods include:

[0073] Step S1, denoising the acquired highway road image to obtain a denoised road image; the denoising process uses a filtering method to filter out medium and high frequency information in the highway road image.

[0074] Step S2, according to the final pixel value of each pixel in the denoised road surface image, determine the fractal feature and fractal dimension eigenvalue of each pixel, and obtain a single feature representation of each pixel based on the fractal feature and fractal dimension eigenvalue; the single feature representations of all pixels constitute a crack feature map.

[0075] Since cracks usually show a high degree of complexity and irregularity, and cracks can appear at different scales, from tiny cracks to larger cracks. These features are difficult to describe using traditional geometric methods, so in this embodiment, fractal features are used to extract crack features. Traditional fractal feature measurement methods can only represent fractal features in the neighborhood, which may miss some important detail textures in the crack image. Therefore, in this embodiment, a method that can represent fractal features at the pixel level is designed to construct a crack feature map.

[0076] Step S3, inputting the crack feature map into the MSF feature extraction network to obtain crack fractal features of different scales.

[0077] Step S4, inputting the crack fractal features of different scales into the feature fusion network for feature fusion to obtain fusion features of different scales.

[0078] Step S5, inputting the fusion features of different scales into the cluster network to obtain the crack detection map of the highway pavement image; the cluster network groups the fusion features of different scales according to the scale size, and uses different convolution kernels to perform convolution operations on the fusion features in different groups to obtain the convolution results; the convolution results corresponding to different groups are added element by element to obtain the crack detection map of the highway pavement image. Grouping by scale size includes: sorting the fusion features of different scales from small to large according to the scale size, grouping the first three fusion features after sorting into one group, and grouping the remaining fusion features into one group.

[0079] The high-speed pavement crack detection method of multi-dimensional feature extraction of this embodiment uses a filtering method to pre-process the pavement image, filter the noise in the image and retain the texture details and boundary information of the crack; uses a fractal feature method to construct a crack feature map to capture the shape and texture features of the crack, and uses an MSF feature extraction network to extract crack features at different scales; constructs a feature fusion network to fuse feature information at different scales, and adds IDS convolution to enhance features and reduce the amount of parameters and calculations; fuses the feature maps of each layer through a cluster network to form a crack detection map, thereby achieving efficient and accurate pavement crack detection. This application has a high crack detection accuracy, strong robustness and good model generalization ability, and has better practical significance.

[0080] In one embodiment, for a highway road image, the gradient change near the noise information corresponds to the intermediate frequency, and the crack texture area usually presents the highest frequency component. In this embodiment, a filter is designed to reduce the intermediate and high frequency signals while maintaining the highest frequency signal. In step S1, the acquired highway road image is denoised to obtain a denoised road image, including:

[0081] Step S11, calculating the filter weight corresponding to each pixel in the highway road image; specifically, the following formula is used:

[0082]

[0083] in, Pixel The corresponding filter weights are represents the Gamma function, is the index of the pixel, is the order, Represents factorial. The value is between 0 and 1. <0.5, it is usually used to extract high-frequency information in the image. Therefore, in this embodiment, in order to extract crack features, The value is 0.3.

[0084] Step S12, according to the filtering weights, weighted sums are performed on the pixel values ​​of the pixels in the neighborhood in four directions around any pixel point to obtain the weighted result of the pixel point in each direction. Specifically, the following formula is used:

[0085]

[0086]

[0087]

[0088]

[0089] in, , , , Represents pixel points ( x,y ) is the weighted result above, below, left, and right, x,y The pixels are x Axis coordinates and y Axis coordinates; is the index of the pixel, is the number of pixels in the neighborhood, Pixel The corresponding filter weights are Represents a pixel ( ), Represents pixel The pixel value of Represents a pixel ( ), Represents pixel The pixel value of .

[0090] In this step, there is a neighborhood in each direction around any pixel point, and the neighborhood refers to the n pixels closest to the currently processed pixel point in a certain direction, where n can be set to 10 or other values ​​without specific limitation. Among them, the n nearest pixels refer to the n pixels closest to the currently processed pixel point in distance.

[0091] Step S13, calculating the average value of the weighted results in all directions as the final pixel value of the pixel point;

[0092] Step S14: the final pixel values ​​of all pixels constitute a denoised road surface image.

[0093] In one embodiment, in step S2, the fractal feature and fractal dimension feature value of each pixel are determined according to the final pixel value of each pixel in the denoised road surface image, including:

[0094] First, determine the fractal dimension eigenvalue of each pixel using the following formula:

[0095]

[0096] in, Pixel The fractal dimension eigenvalue of is a pixel ( x,y ), Pixel The average value of the final pixel value of the pixels in the neighborhood with a radius of r. is the normalization coefficient index, which is used to adjust the scale of the fractal dimension eigenvalue; Indicates absolute value.

[0097] Then, the fractal characteristics of each pixel are determined using the following formula:

[0098]

[0099] in, Pixel The fractal characteristics of Pixel The fractal dimension eigenvalue of Pixel The set of pixels in a neighborhood with a radius of r. express norm, which is used to calculate the difference between two fractal dimension eigenvalues, Indicated in Find the maximum value of the difference among all pixels in .

[0100] In this embodiment, after the fractal feature and the fractal dimension feature value of each pixel are determined, a bitwise XOR operation is performed on the fractal feature and the fractal dimension feature value to obtain a single feature representation of each pixel.

[0101] In one embodiment, Figure 2 shows the structural diagram of the MSF feature extraction network, see Figure 2,The MSF feature extraction network includes 5 feature extraction layers of different scales. The crack fractal features output by the feature extraction layer with a higher scale in two adjacent feature extraction layers are input into the feature extraction layer with a lower scale after downsampling; the 5 feature extraction layers of different scales output a crack fractal feature respectively.

[0102] Here, the MSF feature extraction network includes five layers, each corresponding to a different scale, namely 5×5, 4×4, 3×3, 2×2 and 1×1, and each layer can extract features of a specific scale. Starting from the highest scale 5×5 features, the framework gradually reduces the scale of the features through top-down processing until it reaches the lowest scale 1×1. In this MSF feature extraction network, the 5×5 scale layer can capture the coarsest features, such as overall patterns and shapes. As the level decreases (scale 4×4, scale 3×3, etc.), the captured features gradually transform into finer detail features, such as edges and textures. Therefore, the MSF feature extraction network can form a coarse-to-fine feature representation through top-down processing of multiple layers of different scales.

[0103] In one embodiment, Figure 3 shows a schematic diagram of the structure of the feature fusion network, see Figure 3 ,The feature fusion network includes a convolution module, a feature enhancement module and a fusion module;

[0104] The convolution module is used to perform convolution operations on crack fractal features of different scales to obtain multiple convolution results;

[0105] The feature enhancement module is used to enhance the context information of the lowest-scale crack fractal features to obtain enhanced features;

[0106] The fusion module is used to fuse the enhanced features with multiple convolution results to obtain fused features of different scales.

[0107] In other embodiments, the feature fusion network may also adopt FFN (feed-forward networks).

[0108] Specifically, there are five fractal features of cracks at different scales, denoted as L1, L2, L3, L4, and L5, where the scales of L1, L2, L3, L4, and L5 increase in sequence;

[0109] Enhance the feature of context information of L1 to obtain the enhanced feature C;

[0110] Convolution operations are performed on L1, L2, L3, L4, and L5 respectively to obtain multiple convolution results, which are recorded as M1, M2, M3, M4, and M5;

[0111] C and M1 are added element by element to obtain the fused feature P1;

[0112] P1 and M2 are added element by element to obtain the fused feature P2;

[0113] P2 and M3 are added element by element to obtain the fused feature P3;

[0114] P3 and M4 are added element by element to obtain the fused feature P4;

[0115] P4 and M5 are added element by element to obtain the fused feature P5.

[0116] In this embodiment, starting from the highest scale 5×5 feature L5, the scale of the feature map is gradually reduced by downsampling until the lowest scale 1×1 feature L1 is reached. At each level, the downsampled high-level features are fused with the low-level features by element-by-element addition, and each level outputs a fused feature, denoted as {P1, P2, P3, P4, P5}. In order to enhance the contextual information in the crack feature map, a feature enhancement module is designed in this embodiment to enhance the global and local contextual information obtained by the low-level feature map from the high-level feature map.

[0117] In one embodiment, Figure 4 The structural diagram of the feature enhancement module is shown in Figure 4 , the feature enhancement module is used to:

[0118] Perform convolution operation on the lowest-scale crack fractal features to extract local context information;

[0119] Perform global maximum pooling on the lowest-scale crack fractal features to extract global context information;

[0120] Add the local context information and the global context information element by element to obtain the context information feature map P;

[0121] The context information feature map P is convolved with a convolution rate of 1 and a convolution rate of 3 to obtain two crack feature maps f1 and f2;

[0122] Add the crack feature map f1 and the crack feature map f2 element by element to obtain a new crack feature map F;

[0123] Perform global maximum pooling, convolution, and Sigmoid activation function processing on the new crack feature map F to obtain channel weights;

[0124] The crack feature map f1 and the crack feature map f2 are respectively multiplied element by element with the channel weight to obtain the feature map f1' and the feature map f2' respectively;

[0125] The feature map f1' and the feature map f2' are concatenated to obtain the enhanced feature F'.

[0126] In this embodiment, the enhanced feature F' is used to add M1 element by element, so that the enhanced crack context information can be passed from top to bottom to the low-level feature map in the feature fusion network structure.

[0127] In one embodiment, the high-speed pavement crack detection method using multi-dimensional feature extraction further includes: performing IDS convolution (Intrinsic Dimensional Similarity Convolution) on fused features of different scales respectively to obtain fused features of different scales after feature enhancement.

[0128] Here, the purpose of IDS convolution is to replace the standard convolution with this convolution method, reducing the number of multiplications and additions, thereby reducing the computational complexity and enhancing the network's ability to express crack characteristics.

[0129] In one embodiment, Figure 5 shows a schematic diagram of the cluster network structure, see Figure 5 The cluster network groups the fusion features of different scales according to their scales. For example, P1, P2, and P3 are grouped as G1, and P4 and P5 are grouped as G2. Different convolution kernels are used to perform convolution operations on the fusion features in different groups to obtain convolution results. The convolution results corresponding to different groups are added element by element to obtain the crack detection map of the highway road image.

[0130] Here, a 3×3 convolution kernel is used to perform a convolution operation on P1, P2, and P3 to extract features. This operation helps to retain and enhance local detail information. A 5×5 convolution kernel is applied to P4 and P5 to extract features to capture a wider range of contextual information. The convolution results of the two groups are fused using the element-by-element addition method. This fusion strategy integrates the multi-scale features extracted from G1 and G2 to provide a comprehensive feature representation as the crack detection map output by this embodiment. In other embodiments, a 1×1 convolution kernel can also be used to perform a convolution operation on P1, P2, and P3, a 3×3 convolution kernel can be used to extract features for P4 and P5, or convolution kernels of other sizes can be used, which are not specifically limited here.

[0131] In order to further verify the performance of the high-speed pavement crack detection method based on multi-dimensional feature extraction proposed in this application, three advanced deep learning methods YOLO-v3, FCN-8s, and CrackNet-v were selected for comparative experiments, and the evaluation indicators Recall, Precision, and F1 score were used to compare the detection results. The comparative experimental results are shown in Table 1.

[0132] Table 1

[0133]

[0134] Experiments show that the method of this application has better detection effect on the task of highway pavement crack detection compared with other deep learning methods. It also proves that the method of this application can effectively detect cracks on highway pavements, which has important practical application value for road maintenance and safety assurance.

[0135] Based on the same inventive concept as the high-speed road surface crack detection method using multi-dimensional feature extraction, this embodiment also provides a high-speed road surface crack detection device using multi-dimensional feature extraction corresponding thereto. Figure 6 The structural block diagram of the high-speed pavement crack detection device with multi-dimensional feature extraction is shown, including:

[0136] The denoising module 61 is used to perform denoising on the acquired highway road image to obtain a denoised road image; the denoising process uses a filtering method to filter out medium and high frequency information in the highway road image;

[0137] The crack feature map acquisition module 62 is used to determine the fractal feature and fractal dimension feature value of each pixel point according to the final pixel value of each pixel point in the denoised road surface image, and obtain a single feature representation of each pixel point according to the fractal feature and fractal dimension feature value; the single feature representations of all pixels points constitute a crack feature map;

[0138] A feature extraction module 63 is used to input the crack feature map into the MSF feature extraction network to obtain crack fractal features of different scales;

[0139] A feature fusion module 64 is used to input the crack fractal features of different scales into the feature fusion network for feature fusion to obtain fusion features of different scales;

[0140] The detection module 65 is used to input the fusion features of different scales into the cluster network to obtain the crack detection map of the highway road image; the cluster network groups the fusion features of different scales according to the scale size, and uses different convolution kernels to perform convolution operations on the fusion features in different groups to obtain convolution results; the convolution results corresponding to different groups are added element by element to obtain the crack detection map of the highway road image; grouping according to scale size includes: sorting the fusion features of different scales from small to large according to the scale size, grouping the first 3 fusion features after sorting into one group, and grouping the remaining fusion features into one group.

[0141] The high-speed pavement crack detection device using multi-dimensional feature extraction of this embodiment has the same inventive concept as the high-speed pavement crack detection method using multi-dimensional feature extraction mentioned above. Therefore, the specific implementation method of the device can be seen in the implementation example part of the high-speed pavement crack detection method using multi-dimensional feature extraction mentioned above, and its technical effect corresponds to the technical effect of the above method, which will not be repeated here.

[0142] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A high-speed pavement crack detection method based on multi-dimensional feature extraction, characterized in that: include: De-noising the acquired highway road image to obtain a de-noised road image; The denoising process uses a filtering method to filter out medium and high frequency information in the highway road image; Determine the fractal feature and fractal dimension feature value of each pixel point according to the final pixel value of each pixel point in the denoised road surface image, and obtain a single feature representation of each pixel point according to the fractal feature and fractal dimension feature value; The single feature representation of all pixels constitutes the crack feature map; Inputting the crack feature map into the MSF feature extraction network to obtain crack fractal features of different scales; Inputting the crack fractal features of different scales into a feature fusion network for feature fusion to obtain fusion features of different scales; The fused features of different scales are input into a cluster network to obtain a crack detection map of the highway pavement image; the cluster network groups the fused features of different scales according to the scales, and uses different convolution kernels to perform convolution operations on the fused features in different groups to obtain convolution results; the convolution results corresponding to different groups are added element by element to obtain a crack detection map of the highway pavement image; The grouping according to the scale includes: sorting the fusion features of different scales from small to large according to the scale, grouping the first three fusion features after sorting into one group, and grouping the remaining fusion features into one group; Wherein, according to the final pixel value of each pixel in the denoised road surface image, the fractal feature and the fractal dimension feature value of each pixel are determined, including: Determine the fractal dimension eigenvalue of each pixel using the following formula: in, Pixel The fractal dimension eigenvalue of is a pixel ( x,y ), Pixel The average value of the final pixel value of the pixels in the neighborhood with a radius of r. is the normalized coefficient index; Indicates absolute value; To determine the fractal characteristics of each pixel, the following formula is used: in, Pixel The fractal characteristics of x,y The pixels are x Axis coordinates and y Axis coordinates, Pixel The fractal dimension eigenvalue of Pixel The set of pixels in a neighborhood with a radius of r. express norm; The MSF feature extraction network includes five feature extraction layers of different scales. The crack fractal features output by the feature extraction layer with a higher scale in two adjacent feature extraction layers are input into the feature extraction layer with a lower scale after downsampling; the five feature extraction layers of different scales respectively output a crack fractal feature.

2. The method according to claim 1, characterized in that in, The acquired highway road image is subjected to denoising to obtain a denoised road image, including: Calculate the filter weight corresponding to each pixel in the highway road image; According to the filtering weights, weighted summation is performed on the pixel values ​​of the pixels in the neighborhood in four directions around any pixel point to obtain a weighted result of the pixel point in each direction; Calculate the average value of the weighted results in all directions as the final pixel value of the pixel point; The final pixel values ​​of all pixels constitute the denoised road surface image.

3. The method according to claim 2, characterized in that in, The filter weight corresponding to each pixel in the highway road image is calculated using the following formula: in, Pixel The corresponding filter weights are represents the Gamma function, is the index of the pixel, is the order, Represents factorial.

4. The method according to claim 2, characterized in that in, According to the filtering weights, the pixel values ​​of the pixels in the neighborhood in four directions around any pixel are weighted and summed to obtain the weighted result of the pixel in each direction, using the following formula: in, , , , Represents pixel points ( x,y ) is the weighted result above, below, left, and right, x,y The pixels are x Axis coordinates and y Axis coordinates; is the index of the pixel, is the number of pixels in the neighborhood, Pixel The corresponding filter weights are Represents a pixel ( ), Represents pixel The pixel value of Represents a pixel ( ), Represents pixel The pixel value of .

5. The method according to claim 1, characterized in that The feature fusion network includes a convolution module, a feature enhancement module and a fusion module; The convolution module is used to perform convolution operations on the crack fractal features of different scales respectively to obtain multiple convolution results; The feature enhancement module is used to enhance the feature of the context information of the lowest-scale crack fractal feature to obtain the enhanced feature; The fusion module is used to fuse the enhanced features with the multiple convolution results to obtain fused features of different scales.

6. The method according to claim 5, characterized in that The crack fractal features of different scales include 5, denoted as L1, L2, L3, L4, and L5, wherein the scales of L1, L2, L3, L4, and L5 increase in sequence; Enhance the feature of context information of L1 to obtain the enhanced feature C; Convolution operations are performed on L1, L2, L3, L4, and L5 respectively to obtain multiple convolution results, which are recorded as M1, M2, M3, M4, and M5; C and M1 are added element by element to obtain the fused feature P1; P1 and M2 are added element by element to obtain the fused feature P2; P2 and M3 are added element by element to obtain the fused feature P3; P3 and M4 are added element by element to obtain the fused feature P4; P4 and M5 are added element by element to obtain the fused feature P5.

7. The method according to claim 5, characterized in that The feature enhancement module is used to: Performing a convolution operation on the crack fractal feature of the lowest scale to extract local context information; Performing global maximum pooling on the crack fractal features of the lowest scale to extract global context information; Add the local context information and the global context information element by element to obtain a context information feature map P; Perform convolution operations with a convolution rate of 1 and a convolution rate of 3 on the context information feature map P, respectively, to obtain two crack feature maps f1 and f2; Add the crack feature map f1 and the crack feature map f2 element by element to obtain a new crack feature map F; Performing global maximum pooling operation, convolution operation, and Sigmoid activation function processing on the new crack feature map F to obtain channel weights; The crack feature map f1 and the crack feature map f2 are respectively multiplied element by element with the channel weight to obtain a feature map f1' and a feature map f2' respectively; The feature map f1' and the feature map f2' are concatenated to obtain an enhanced feature F'.

8. The method according to claim 1, characterized in that The method further comprises: IDS convolution is performed on the fusion features of different scales respectively to obtain the fusion features of different scales after feature enhancement.

9. A high-speed pavement crack detection device with multi-dimensional feature extraction, characterized in that: include: A denoising module is used to perform denoising on the acquired highway road image to obtain a denoised road image; The denoising process uses a filtering method to filter out medium and high frequency information in the highway road image; A crack feature map acquisition module is used to determine the fractal feature and fractal dimension feature value of each pixel point according to the final pixel value of each pixel point in the denoised road surface image, and obtain a single feature representation of each pixel point according to the fractal feature and fractal dimension feature value; the single feature representations of all pixels points constitute a crack feature map; A feature extraction module, used for inputting the crack feature map into the MSF feature extraction network to obtain crack fractal features of different scales; A feature fusion module, used for inputting the crack fractal features of different scales into a feature fusion network for feature fusion to obtain fusion features of different scales; A detection module, for inputting the fused features of different scales into a cluster network to obtain a crack detection map of the highway pavement image; the cluster network groups the fused features of different scales according to the scales, and uses different convolution kernels to perform convolution operations on the fused features in different groups to obtain convolution results; the convolution results corresponding to different groups are added element by element to obtain a crack detection map of the highway pavement image; the grouping according to scale includes: sorting the fused features of different scales from small to large according to the scale, grouping the first three fused features after sorting into one group, and grouping the remaining fused features into one group; Wherein, according to the final pixel value of each pixel in the denoised road surface image, the fractal feature and the fractal dimension feature value of each pixel are determined, including: Determine the fractal dimension eigenvalue of each pixel using the following formula: in, Pixel The fractal dimension eigenvalue of is a pixel ( x,y ), Pixel The average value of the final pixel value of the pixels in the neighborhood with a radius of r. is the normalized coefficient index; Indicates absolute value; To determine the fractal characteristics of each pixel, the following formula is used: in, Pixel The fractal characteristics of x,y The pixels are x Axis coordinates and y Axis coordinates, Pixel The fractal dimension eigenvalue of Pixel The set of pixels in a neighborhood with a radius of r. express norm; The MSF feature extraction network includes five feature extraction layers of different scales. The crack fractal features output by the feature extraction layer with a higher scale in two adjacent feature extraction layers are input into the feature extraction layer with a lower scale after downsampling; the five feature extraction layers of different scales respectively output a crack fractal feature.

Images