Highway pavement micro crack detection system based on improved group algorithm

Abstract

The application discloses a highway pavement micro-crack detection system based on an improved group algorithm, which comprises an image acquisition module, an image preprocessing module, a parameter optimization module, an image segmentation module, a feature extraction module, a region reconstruction module and a detection and output module. The image acquisition module is used for outputting a standardized highway pavement image. The image preprocessing module is used for generating a pretreated highway pavement image. The parameter optimization module performs global search and local refinement search on the pretreated highway pavement image based on a double-stage bitterfish optimization algorithm model. The image segmentation module generates a preliminary micro-crack candidate region image. The feature extraction module forms a deep feature map. The region reconstruction module outputs a final highway pavement micro-crack region segmentation result. The detection and output module performs candidate region detection and positioning on the highway pavement micro-crack region segmentation result, and outputs a final highway pavement micro-crack detection result containing micro-crack position and region information. The application improves the response capability to weak edges, and in the local refinement stage, a crack topological structure consistency function is fused, so that the detection structure restoration precision of the micro-crack region is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crack detection, and in particular to a highway pavement micro-crack detection system based on an improved group algorithm. BACKGROUND

[0002] With the acceleration of urbanization and the improvement of transportation intensity, highway pavement is prone to early micro-crack disease during long-term high-load use, which is difficult to detect at the initial stage, but if not found and treated in time, it is easy to evolve into more serious structural damage, thereby affecting traffic safety and road life. Therefore, developing high-precision and high-efficiency highway pavement micro-crack detection technology has become a key task in the field of road maintenance and management.

[0003] At present, the detection methods of highway pavement micro-cracks mainly include manual inspection, laser scanning, infrared thermal imaging and image processing technology, among which the manual inspection method is limited by subjective judgment and is not suitable for large-scale rapid detection. Although laser scanning and thermal imaging have certain automation capabilities, they have high equipment costs, complex operation, and limited resolution when detecting fine cracks, and there are problems of missed detection or false detection. Therefore, the image processing method based on computer vision has gradually become a research hotspot.

[0004] However, the existing image processing detection methods usually rely on fixed threshold segmentation, edge detection or shallow feature extraction, and it is difficult to deal with the low contrast, weak edge and irregular shape problems of micro-crack images in complex road scenes. For example, the traditional Canny or Sobel operator is prone to produce false edges when dealing with random noise or uneven illumination in the image, resulting in inaccurate crack positioning. The fixed threshold strategy is difficult to adapt to the diversity of different image quality and crack shape. In addition, most image segmentation methods lack analysis means for the connectivity and topological consistency of micro-crack structure, and are prone to appear broken, adhesion or missing area, which affects the integrity and accuracy of the final detection result.

[0005] Therefore, there is an urgent need for a new micro-crack detection method that combines intelligent optimization algorithm and deep image feature modeling capability to effectively improve the stability, accuracy and practicality of detection. SUMMARY

[0006] One object of the present application is to provide a highway pavement micro-crack detection system based on an improved group algorithm, which improves the response capability to weak edges and integrates crack topological structure consistency function in the local refinement stage, thereby improving the detection structure restoration precision of micro-crack regions.

[0007] According to the highway pavement micro-crack detection system based on the improved group algorithm of the embodiment of the present application, the system comprises:

[0008] An image acquisition module is configured to acquire a road surface image and complete standardization processing, and output a standardized road surface image.

[0009] An image preprocessing module is configured to perform noise suppression and edge enhancement operations on the standardized road surface image, and generate a preprocessed road surface image.

[0010] A parameter optimization module is configured to perform global search and local refinement search on the preprocessed road surface image based on a two-stage bitterfish optimization algorithm model, and output final fine image segmentation parameters.

[0011] An image segmentation module is configured to take the preprocessed road surface image and the final fine image segmentation parameters as inputs, perform preliminary micro-crack region image segmentation, and generate a preliminary micro-crack candidate region map.

[0012] A feature extraction module is configured to input the preliminary micro-crack candidate region map into a multi-scale hollow convolution module, generate a multi-scale feature response map, and perform channel attention mechanism and feature splicing processing to form a deep feature map.

[0013] A region reconstruction module is configured to perform adaptive threshold binarization and morphological reconstruction on the deep feature map, and output a final road surface micro-crack region segmentation result.

[0014] A detection and output module is configured to perform candidate region detection and positioning on the road surface micro-crack region segmentation result, judge and screen the road surface micro-crack region using region contour analysis, and output a final road surface micro-crack detection result containing micro-crack position and region information.

[0015] A road surface micro-crack detection method based on an improved group algorithm is applied to a road surface micro-crack detection system based on an improved group algorithm, and includes the following steps:

[0016] S1. Collect road surface images and perform standard processing to form standardized road surface images.

[0017] S2. Preprocess the standardized road surface images to generate preprocessed road surface images.

[0018] S3. Construct a two-stage bitterfish optimization algorithm model based on the preprocessed road surface images, initialize the optimization population in the global search stage, and evaluate the image segmentation parameters according to the fitness function to obtain coarse-grained image segmentation parameters.

[0019] S4. Enter the local refinement search stage of the two-stage bitterfish optimization algorithm model using the coarse-grained image segmentation parameters, and obtain final fine image segmentation parameters through local disturbance and regulation.

[0020] S5. Taking the pre-processed highway pavement image and the final fine image segmentation parameters as inputs, performing an image segmentation operation to generate a preliminary micro-crack candidate region map;

[0021] S6. Inputting the preliminary micro-crack candidate region map into a multi-scale hollow convolution module to generate a multi-scale feature response map, and performing channel attention mechanism and feature splicing processing to form a deep feature map;

[0022] S7. Performing adaptive threshold binarization processing and morphological reconstruction on the deep feature map to eliminate residual noise and repair crack continuity, and outputting the final highway pavement micro-crack region segmentation result;

[0023] S8. Performing candidate region detection and positioning on the highway pavement micro-crack region segmentation result, using region contour analysis to judge and screen the highway pavement micro-crack region, and outputting the final highway pavement micro-crack detection result containing the micro-crack position and region information.

[0024] Optionally, the S1 comprises the following steps:

[0025] S11. Setting a highway pavement image acquisition frequency, acquiring a highway pavement image through a highway pavement image acquisition device, and the highway pavement image acquisition resolution is R w ×R h , wherein R w represents the width of the highway pavement image, and R h represents the height of the highway pavement image;

[0026] S12. Performing size normalization processing on the acquired original highway pavement image to uniformly scale different size highway pavement images to a standard size R s = W s × H s , wherein R s represents the standard highway pavement image size, W s and H s are the uniform width and height of the standard highway pavement image, respectively;

[0027] S13. Performing color space conversion operation on the size-normalized highway pavement image to convert the RGB highway pavement image into a gray space highway pavement image;

[0028] S14. Implementing highway pavement image enhancement processing on the gray highway pavement image to improve the distinguishability of micro-cracks in low-contrast regions, obtaining an enhanced highway pavement image I enh , and taking the enhanced highway pavement image as the standardized highway pavement image.

[0029] Optionally, the S2 comprises the following steps:

[0030] S21. Based on the standardized highway pavement image I enh , a Gaussian filter is used for preliminary noise suppression processing of the highway pavement image, to suppress random noise generated in the environmental acquisition process, and a Gaussian smoothed highway pavement image I gauss is generated;

[0031] S22. The Gaussian smoothed highway pavement image is filtered using the gradient direction of the highway pavement image and the brightness difference between adjacent pixels to regulate the filtering result, and an edge-preserving highway pavement image is generated;

[0032] S23. The edge-preserving highway pavement image is processed by multi-channel filtering, and different directional guide filtering channels {I θ} are constructed, where θ ∈ {0°, 45°, 90°, 135°}, each directional filtering channel enhances the micro-crack linear structure through directional response, and a set of directional enhanced highway pavement images {I dir,θ} is generated;

[0033] S24. The multi-channel directional enhanced highway pavement image set {I dir,θ} is processed by response fusion, and a weighted superposition strategy is used to calculate the fusion highway pavement image I fused ;

[0034] S25. The edge detection processing is performed on the fusion highway pavement image, and the Sobel operator is used to calculate the gradient response values G x (x,y) and G y (x,y) of the highway pavement image in the horizontal and vertical directions, and the edge response map I edge of the highway pavement image is obtained;

[0035] S26. The edge response map I edge is fused with the Gaussian smoothed highway pavement image I gauss at the pixel level to generate a preprocessed highway pavement image I pre containing noise suppression and edge enhancement features:

[0036]

[0037] where θ1 is a noise modulation factor for enhancing the noise suppression effect of the Gaussian smoothed highway pavement image I gauss , θ2 is an attenuation constant for regulating the attenuation rate of the exponential function on the Gaussian smoothed highway pavement image I gauss in different brightness regions, and θ3 is an edge enhancement scaling factor for amplifying the edge response map I edgeThe contribution of the edge detail preserving micro-crack in the fusion process, θ4 is an edge sensitivity adjustment factor, the edge response value is adjusted by the hyperbolic tangent function to distinguish the low response and high response area, μ is the edge threshold parameter, set the reference threshold of edge response, only when the edge response map I edge (x,y) exceeds the edge threshold parameter, the edge information is enhanced, exp(·) and tanh(·) represent the exponential function and hyperbolic tangent function respectively.

[0038] Optionally, the S3 comprises the following steps:

[0039] S31. Construct a two-stage bitterfish optimization algorithm model including a global search stage and a local refinement search stage, initialize the bitterfish optimization population Bitterfish optimization individual X i Indicates a set of preprocessed highway pavement image segmentation parameters, including the binary threshold T i , filter kernel size K i , and feature enhancement weight parameter λ i , wherein N is the size of the bitterfish optimization population;

[0040] S32. For the detection characteristics of highway pavement micro-crack area segmentation, based on the bitterfish optimization individual X i Construct fitness function F(X i ):

[0041]

[0042] Wherein, is the inter-class variance between the micro-crack area and the background area in the preprocessed highway pavement image under the action of the preprocessed highway pavement image segmentation parameter X i , which measures the segmentation discrimination degree of the micro-crack area, E edge (X i ) is the edge saliency of the preprocessed highway pavement image under the action of the preprocessed highway pavement image segmentation parameter X i , which is used to evaluate the clarity of the micro-crack edge, R conn (X i ) is the connectivity index of the micro-crack area in the preprocessed highway pavement image under the action of the preprocessed highway pavement image segmentation parameter X i , which is used to evaluate the continuity of the micro-crack area, α1, α2, α3 are the corresponding weight coefficients;

[0043] S33. In the global search stage of the two-stage bitterfish optimization, the current optimal bitterfish optimization individual position i and the average position of the bitterfish optimization population are calculated according to the fitness function F(X ​and a crack sensitivity factor is introduced

[0044]

[0045] wherein, is the edge saliency of the current optimal Kuhsh optimization individual corresponding to the image, is the average value of the edge saliency of all Kuhsh optimization individuals in the current Kuhsh optimization population, and β is a crack sensitivity adjustment coefficient;

[0046] S34. Using the crack sensitivity factor the position of the Kuhsh optimization individual is updated:

[0047]

[0048] wherein, respectively represent the positions of the Kuhsh optimization individual at the t-th and t+1-th iterations, and r2 and r1 are random factors;

[0049] S35. The crack structure convergence threshold of the global search stage is set, and the fitness value of the current optimal Kuhsh optimization individual is calculated according to the fitness function. When the fitness value of the current optimal Kuhsh optimization individual is greater than or equal to the crack structure convergence threshold, the iteration of the global search stage is terminated, and the coarse-grained image segmentation parameter X used for preliminary segmentation of the micro-cracks of the highway pavement is obtained global , otherwise, return to step S33 to continue the global search process.

[0050] Optionally, the S4 comprises the following steps:

[0051] S41. In the local refinement search stage of the two-stage Kuhsh optimization algorithm model, a set of local Kuhsh optimization individuals X' is constructed wherein M << N, and each local Kuhsh optimization individual X' i is added with a disturbance vector Δ i generated by the coarse-grained image segmentation parameter:

[0052]

[0053] wherein, represents a multi-dimensional Gaussian disturbance distribution with a mean of 0 and a covariance matrix Σ loc , which is used for fine-grained sampling of the local search space;

[0054] S42. Based on the preprocessed highway pavement image I pre , a micro-crack topology graph G=(V, E) is constructed, wherein V represents a set of crack candidate points extracted by edge detection, and E represents a set of crack structure edges connected by local gradient direction and distance threshold, and a crack region consistency function C(X') is constructed in combination with the topological structurei (,G), used to evaluate the degree of matching between the current microcrack segmentation result and the actual microcrack topology:

[0055]

[0056] Wherein, R(X′) i ) indicates that in parameter X′ i The microcrack segmentation region obtained below, θ uv The angle representing the consistency of the direction of the crack structure edge (u,v) is given. This is an indicator function used to determine whether the endpoints of the crack structure are simultaneously in the candidate region;

[0057] S43. Based on topological structure, a consistency function for the crack region is constructed to build a composite fitness function to evaluate the local bitter fish optimization individual:

[0058] F local (X′ i )=γ1·F(X′ i )+γ2·C(X′ i ,G);

[0059] Wherein, F(X′) i ) represents the fitness function of the individual when applying local bitter fish optimization, and γ1 and γ2 are the weight coefficients of the global search stage and the local refinement search stage.

[0060] S44. Optimize individual X′ for local bitter fish. i The strategy of optimizing and updating bitter fish is applied, and the update direction of the local search is set to the current locally optimal bitter fish individual X′. best Optimize the individual position X′ of the current bitter fish i The vector difference between them is used, and a dynamic step size factor is introduced to finely control the search process. The dynamic step size factor is based on the composite fitness function F of the local bitter fish optimization individual. local (X′ i The composite fitness function F with the preset target target Dynamic decay adjustment is implemented, with the step size gradually decreasing as individual fitness improves;

[0061] S45. When the termination condition is met—the fitness value of the currently local optimal bitter fish individual is greater than or equal to the local crack structure convergence threshold—output the final fine image segmentation parameters X. final ={T final ,K final ,λ final And complete the local search for the two-stage bitter fish optimization.

[0062] Optionally, S5 includes the following steps:

[0063] S51. Based on the pre-processed road surface image I pre and the final fine image segmentation parameter X final , a preliminary image segmentation operation of the road surface micro-cracks is performed to generate a preliminary segmentation image I seg :

[0064]

[0065] where I seg (x, y) represents the pixel value at coordinate (x, y) after the preliminary image segmentation, and T final is the optimal binary threshold in the final fine image segmentation parameter, used to distinguish the micro-crack region from the non-crack region.

[0066] S52. Based on the filter kernel size K final in the final fine image segmentation parameter, a morphological closing operation is performed on the preliminary segmentation image I seg to generate a connectivity-enhanced preliminary micro-crack candidate region map I crack :

[0067]

[0068] where and represent the morphological dilation and erosion operations, respectively, and the size of the structural element is determined by the final fine image segmentation parameter K final , used to smooth the interference and broken regions in the micro-crack candidate region map.

[0069] S53. Based on the feature enhancement weight parameter λ final in the final fine image segmentation parameter, a micro-crack region feature enhancement operation is performed on the preliminary micro-crack candidate region map I crack to obtain an enhanced preliminary micro-crack candidate region map I 增强 :

[0070] I 增强 (x, y) = I crack (x, y) · [1 + λ final · I edge (x, y)];

[0071] where λ final is the weight coefficient, used to enhance the edge features of the micro-crack region.

[0072] Optionally, the S6 comprises the following steps:

[0073] S61. The enhanced preliminary micro-crack candidate region map I 增强Input a multi-scale dilated convolution submodule, in which a set of dilation rates D = {d1, d2, d3, d4} is defined, and four sets of dilated convolution branches are constructed in parallel, each set of branches using a corresponding dilation rate d. i Hollow convolution kernel Perform a dilated convolution operation on the input image to generate a dilated feature map. :

[0074]

[0075] in, This indicates the void ratio d i Hole feature response at position (x,y);

[0076] S62. Map all void features The input channel attention submodule assigns attention weights to the multi-scale channel features extracted by different convolutional branches, resulting in a channel-weighted response map.

[0077]

[0078] Where Pool(·) represents a combination of global average pooling and max pooling, and MLP(·) is an inter-channel weight learning network. The void ratio is d. i Feature map response weights;

[0079] S63. Channel-weighted response diagram The input feature fusion decoding submodule performs channel-dimensional concatenation to obtain the fused feature map F. fused Then fuse the feature map F fused The input is an upsampling decoding network, which performs layer-by-layer upsampling and spatial feature reconstruction using deconvolution operations to generate a depth feature map F with the same size as the input image. deep .

[0080] Optionally, S7 includes the following steps:

[0081] S71. Transfer the depth feature map F deep The input adaptive threshold module automatically generates a spatial dynamic threshold map T based on local statistical features. adapt (x,y) is combined with the pixel response values ​​to perform a binarization operation, generating the preliminary segmentation result image I. bin ;

[0082] S72. Preliminary segmentation result Figure I bin Perform morphological reconstruction operations, utilizing the fine linearity and connectivity characteristics of microcracks, and employing the structuring element S. recPerforming morphological closing operation and thinning processing to generate structure-continuous micro-crack region map I recon :

[0083]

[0084] wherein, are expansion and corrosion operations respectively, Thin(·) represents thinning operation, and the size of structure element S rec is set according to the width range of target crack;

[0085] S73. After structure reconstruction, micro-crack region map I recon is executed to perform connected domain screening and area constraint processing to eliminate isolated small regions with area less than a set threshold A min , to form a final micro-crack region segmentation map I final .

[0086] Optionally, the S8 comprises the following steps:

[0087] S81. For the final micro-crack region segmentation map I final , perform connected domain extraction operation to obtain the boundary contour set of all independent micro-crack candidate regions , and calculate the morphological parameter set of each candidate region respectively wherein, A k is the area of the kth region, is the aspect ratio, is the average response value of the internal depth feature of the region;

[0088] S82. Combine the final fine image segmentation parameters to construct dynamic micro-crack region classification rules, and set the following threshold values:

[0089] Area dynamic threshold wherein, α1 is the area adjustment coefficient;

[0090] Low response intensity threshold T low = α2·T final , high threshold T high = α3·λ final ·T final , wherein α2, α3 are response adjustment coefficients respectively;

[0091] S83. According to the performance characteristics of micro-crack morphology in actual highway detection, the detection results are divided into three categories: suspected micro-crack region, obvious micro-crack region and non-crack region:

[0092] The non-crack region classification rule is to satisfy any one of the following conditions: A k <A thresh , or and

[0093] Suspected micro crack region classification rule is to meet all the following conditions: A k ≥ A thresh , And

[0094] The obvious micro crack region classification rule is to meet all the following conditions: A k ≥ A thresh , And

[0095] S84. The position information extraction and structured output of the classified crack candidate region is carried out, the boundary center point coordinates, the circumscribed rectangle boundary box and the region number, the classification label and the morphological parameters of each obvious micro crack region are extracted, and the final highway pavement micro crack detection result set is constituted.

[0096] The beneficial effects of the present application are:

[0097] (1) The present application introduces a double-stage bitter fish optimization algorithm model, introduces a double-layer mechanism of global search and local refinement search in the optimization process of image segmentation parameters, realizes a gradual optimization strategy from coarse granularity to refinement, and optimizes the model. The crack sensitivity factor is introduced in the global search stage to improve the response ability of the weak edge, and the crack topological structure consistency function is fused in the local refinement stage, which effectively improves the structure restoration accuracy of the micro crack region.

[0098] (2) The deep feature encoding network combined with multi-scale hollow convolution and channel attention mechanism introduced in the micro crack feature extraction process can more comprehensively capture crack texture information of different sizes, directions and densities. By setting a set of hole rates to construct multiple parallel hollow convolution branches, the key crack region features are strengthened by combining the response weight distribution mechanism between channels, which has a significant advantage in weak contrast and complex background texture road surface images. BRIEF DESCRIPTION OF DRAWINGS

[0099] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation of the present application. In the drawings:

[0100] Figure 1 A flow chart of a highway pavement micro crack detection system based on an improved group algorithm is provided. DETAILED DESCRIPTION

[0101] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0102] Reference Figure 1 A highway pavement micro-crack detection system based on an improved group algorithm, comprising:

[0103] An image acquisition module for acquiring highway pavement images and completing standardization processing, and outputting standardized highway pavement images;

[0104] An image preprocessing module for performing noise suppression and edge enhancement operations on the standardized highway pavement images to generate preprocessed highway pavement images;

[0105] A parameter optimization module for performing global search and local refinement search on the preprocessed highway pavement images based on a two-stage bitterfish optimization algorithm model, and outputting final fine image segmentation parameters;

[0106] An image segmentation module for performing preliminary micro-crack region image segmentation on the preprocessed highway pavement images and the final fine image segmentation parameters as inputs, and generating a preliminary micro-crack candidate region map;

[0107] A feature extraction module for inputting the preliminary micro-crack candidate region map into a multi-scale hollow convolution module to generate multi-scale feature response maps, and performing channel attention mechanism and feature splicing processing to form a deep feature map;

[0108] A region reconstruction module for performing adaptive threshold binarization and morphological reconstruction on the deep feature map, and outputting final highway pavement micro-crack region segmentation results;

[0109] A detection and output module for performing candidate region detection and positioning on the highway pavement micro-crack region segmentation results, and using region contour analysis to judge and screen the highway pavement micro-crack regions, and outputting final highway pavement micro-crack detection results containing micro-crack position and region information.

[0110] A highway pavement micro-crack detection method based on an improved group algorithm, applied to a highway pavement micro-crack detection system based on an improved group algorithm, comprising the following steps:

[0111] S1. Collecting highway pavement images and performing standard processing to form standardized highway pavement images;

[0112] S2. Preprocessing the standardized highway pavement images to generate preprocessed highway pavement images;

[0113] S3. Utilize the two-stage bitter fish optimization algorithm model based on the pre-processed highway pavement image to initialize the optimization population in the global search stage, and evaluate the image segmentation parameters according to the fitness function to obtain the coarse-grained image segmentation parameters;

[0114] S4. Use the coarse-grained image segmentation parameters to enter the local refinement search stage of the two-stage bitter fish optimization algorithm model, and obtain the final fine image segmentation parameters through local disturbance and regulation;

[0115] S5. Take the pre-processed highway pavement image and the final fine image segmentation parameters as input, perform image segmentation operation, and generate a preliminary micro-crack candidate region map;

[0116] S6. Input the preliminary micro-crack candidate region map into the multi-scale hollow convolution module to generate a multi-scale feature response map, and perform channel attention mechanism and feature splicing processing to form a deep feature map;

[0117] S7. Perform adaptive threshold binarization processing and morphological reconstruction on the deep feature map to eliminate residual noise and repair crack continuity, and output the final highway pavement micro-crack region segmentation result;

[0118] S8. Perform candidate region detection and positioning on the highway pavement micro-crack region segmentation result, use region contour analysis to judge and screen the highway pavement micro-crack region, and output the final highway pavement micro-crack detection result containing the micro-crack position and region information.

[0119] In this embodiment, S1 includes the following steps:

[0120] S11. Set the highway pavement image acquisition frequency, acquire the highway pavement image through the highway pavement image acquisition device, and the highway pavement image acquisition resolution is R w ×R h , wherein R w represents the width of the highway pavement image, and R h represents the height of the highway pavement image;

[0121] S12. Perform size normalization processing on the collected original highway pavement image, and uniformly scale different size highway pavement images to a standard size R s = W s × H s , wherein R s represents the standard highway pavement image size, W s and H s are the uniform width and height of the standard highway pavement image, respectively;

[0122] S13. Perform a color space conversion operation on the size-normalized road surface image to convert the RGB road surface image to a grayscale space road surface image;

[0123] S14. Perform a road surface image enhancement process on the grayscale road surface image to improve the distinguishability of micro-cracks in low-contrast regions, to obtain an enhanced road surface image I enh , and take the enhanced road surface image as the standardized road surface image.

[0124] In this embodiment, S2 includes the following steps:

[0125] S21. Based on the standardized road surface image I enh , a Gaussian filter is used to perform a preliminary noise suppression process on the road surface image to suppress random noise generated during environmental acquisition, to generate a Gaussian-smoothed road surface image I gauss ;

[0126] S22. The Gaussian-smoothed road surface image is filtered using the brightness difference between the gradient direction of the road surface image and the adjacent pixels to regulate the filtering result, to generate an edge-preserving road surface image;

[0127] S23. The edge-preserving road surface image is subjected to multi-channel filtering processing, and different directional guide filtering channels {I θ} are constructed, where θ ∈ {0°, 45°, 90°, 135°}, and each directional filtering channel enhances the micro-crack linear structure through directional response, and generates a directional enhanced road surface image set {I dir,θ};

[0128] S24. The multi-channel directional enhanced road surface image set {I dir,θ} is subjected to response fusion processing, and a weighted superposition strategy is used to calculate a fused road surface image I fused ;

[0129] S25. The fused road surface image is subjected to edge detection processing, and a Sobel operator is used to calculate the gradient response values G x (x,y) and G y (x,y) of the road surface image in the horizontal and vertical directions, to obtain an edge response map I edge of the road surface image;

[0130] S26. The edge response map I edge is subjected to pixel-level fusion processing with the Gaussian-smoothed road surface image I gauss , to generate a preprocessed road surface image I pre containing noise suppression and edge enhancement features:

[0131]

[0132] where θ1 is a noise modulation factor for enhancing the noise suppression effect of the Gaussian-smoothed road surface image I gauss , θ2 is a decay constant for regulating the decay rate of the exponential function in different brightness regions affecting the Gaussian-smoothed road surface image I gauss , θ3 is an edge enhancement scaling factor for amplifying the contribution of the edge response map I edge to preserve the edge details of micro-cracks in the fusion process, and θ4 is an edge sensitivity adjustment factor for adjusting the non-linear mapping of the edge response value through the hyperbolic tangent function to distinguish low-response and high-response regions. μ is an edge threshold parameter that sets the reference threshold for the edge response. Only when the edge response map I edge (x, y) exceeds the edge threshold parameter, the edge information is enhanced. exp(·) and tanh(·) represent the exponential function and the hyperbolic tangent function, respectively.

[0133] In this embodiment, S3 includes the following steps:

[0134] S31. Construct a two-stage grouperfish optimization algorithm model including a global search stage and a local refinement search stage, and initialize the grouperfish optimization population X i represents a set of pre-processed road surface image segmentation parameters, including a binary threshold T i , a filter kernel size K i , and a feature enhancement weight parameter λ i , where N is the size of the grouperfish optimization population.

[0135] S32. For the detection characteristics of the micro-crack region segmentation of the road surface, based on the grouperfish optimization individual X i , construct the fitness function F(X i ):

[0136]

[0137] where is the inter-class variance between the micro-crack region and the background region in the pre-processed road surface image under the action of the pre-processed road surface image segmentation parameter X i , which measures the segmentation discrimination of the micro-crack region, E edge (X i ) is the edge saliency of the pre-processed road surface image under the action of the pre-processed road surface image segmentation parameter X i , which is used to evaluate the clarity of the micro-crack edge, and R conn (X i ) is the edge saliency of the pre-processed road surface image under the action of the pre-processed road surface image segmentation parameter X i .The connectivity index of microcrack regions in preprocessed highway pavement images under the action of the function is used to evaluate the continuity of microcrack regions. α1, α2, and α3 are the corresponding weight coefficients.

[0138] S33. In the global search phase of the two-stage bitter fish optimization, according to the fitness function F(X) i Calculate the current optimal position of the bitter fish. and the average position of the optimized population of bitter fish And introduce a crack sensitivity factor

[0139]

[0140] in, To optimize the image edge salience of the current best bitter fish individual, β is the average edge salience of all optimized individuals of bitter fish in the current optimized population, and β is the crack sensitivity adjustment coefficient;

[0141] S34. Utilizing crack sensitivity factors Update the individual positions of the bitter fish:

[0142]

[0143] in, Let r1 and r2 represent the positions of the optimized individual of the bitter fish at the t and t+1th iterations, respectively, where r2 and r1 are random factors;

[0144] S35. Set the crack structure convergence threshold for the global search phase, and calculate the fitness value of the current optimal individual based on the fitness function. When the fitness value of the current optimal individual is greater than or equal to the crack structure convergence threshold, terminate the iteration of the global search phase and obtain the coarse-grained image segmentation parameters X for the initial segmentation of microcracks in highway pavement. global Otherwise, return to step S33 to continue the global search process.

[0145] In this embodiment, S4 includes the following steps:

[0146] S41. In the local refinement search phase of the two-stage bitter fish optimization algorithm model, construct a set of locally optimized bitter fish individuals. Where M << N, and each local bitter fish optimization individual X′ i Add perturbation vector Δ to coarse-grained image segmentation parameters i generate:

[0147]

[0148] in, This indicates that the mean is 0 and the covariance matrix is ​​Σ.loc a multi-dimensional Gaussian perturbation distribution for fine-grained sampling of the local search space;

[0149] S42. Based on the pre-processed highway pavement image I pre Construct a micro-crack topology graph G=(V, E), where V represents a set of crack candidate points extracted by edge detection, and E represents a set of crack structure edges connected by local gradient direction and distance threshold, and a crack region consistency function C(X′ i , G) is constructed based on the topology structure, which is used to evaluate the matching degree of the current micro-crack segmentation result and the real topology structure of the micro-crack:

[0150]

[0151] Where R(X′ i ) represents the micro-crack segmentation region obtained under the parameter X′ i , θ uv is the direction consistency angle of the crack structure edge (u, v), is an indicator function used to determine whether the endpoints of the crack structure edge are in the candidate region at the same time;

[0152] S43. Based on the topology structure, a crack region consistency function is constructed to evaluate the local fish optimization individual:

[0153] F local (X′ i ) = γ1·F(X′ i ) + γ2·C(X′ i , G);

[0154] Where F(X′ i ) is the fitness function of the local fish optimization individual, and γ1 and γ2 are the weight coefficients of the global search stage and the local refinement search stage;

[0155] S44. The local fish optimization individual X′ i is updated by the fish optimization update strategy, and the update direction of the local search is set as the vector difference between the current local optimal fish optimization individual X′ best and the current fish optimization individual position X′ i , and a dynamic step factor is introduced to fine control the search process, and the dynamic step factor is dynamically attenuated and adjusted according to the composite fitness function F local (X′ i ) of the local fish optimization individual and the preset target composite fitness function F target , and the step gradually decreases with the improvement of the individual fitness;

[0156] S45. When the termination condition is met, i.e., the fitness value of the current locally optimal fish swarm optimization individual is greater than or equal to the local crack structure convergence threshold, output the final fine image segmentation parameter X final = {T final , K final , λ final} and complete the local search of the two-stage fish swarm optimization.

[0157] In this embodiment, S5 includes the following steps:

[0158] S51. Based on the pre-processed highway pavement image I pre and the final fine image segmentation parameter X final , perform a preliminary image segmentation operation of the highway pavement micro-cracks to generate a preliminary segmentation image I seg :

[0159]

[0160] where I seg (x, y) represents the pixel value at coordinate (x, y) after preliminary image segmentation, T final is the optimal binary threshold in the final fine image segmentation parameter, used to distinguish the micro-crack region and the non-crack region;

[0161] S52. Based on the filter kernel size K final in the final fine image segmentation parameter, perform a morphological closing operation on the preliminary segmentation image I seg to generate a connectivity-enhanced preliminary micro-crack candidate region map I crack :

[0162]

[0163] where and represent morphological dilation and erosion operations, respectively, and the size of the structural element is determined by the final fine image segmentation parameter K final , used to smooth the interference and broken areas in the micro-crack candidate region map;

[0164] S53. Based on the feature enhancement weight parameter λ final in the final fine image segmentation parameter, perform a micro-crack region feature enhancement operation on the preliminary micro-crack candidate region map I crack to obtain an enhanced preliminary micro-crack candidate region map I 增强 :

[0165] I 增强 (x, y) = I crack (x, y) · [1 + λ final · I edge(x, y)];

[0166] where λ final is a weight coefficient for enhancing the edge feature of the micro crack region.

[0167] In this embodiment, S6 includes the following steps:

[0168] S61. input the enhanced preliminary micro crack candidate region map I 增强 into a multi-scale dilated convolution submodule, set a set of dilated rates D = {d1, d2, d3, d4} in the multi-scale dilated convolution submodule, and construct four groups of dilated convolution branches in parallel, each group of branches uses a dilated convolution kernel corresponding to the dilated rate d i . Perform a dilated convolution operation on the input image to generate a dilated feature map

[0169]

[0170] wherein, represents the dilated feature response at position (x, y) under the dilated rate d i .

[0171] S62. input all dilated feature maps into a channel attention submodule, and perform attention weight distribution on the multi-scale channel features extracted by different convolution branches to obtain a channel weighted response map

[0172]

[0173] wherein, Pool(·) represents a global average pooling combined with maximum pooling operation, and MLP(·) is an inter-channel weight learning network, represents the feature response weight of the feature map with a dilated rate of d i .

[0174] S63. input the channel weighted response map into a feature fusion decoding submodule, perform channel dimension splicing to obtain a fusion feature map F fused , and then input the fusion feature map F fused into an up-sampling decoding network to perform layer-by-layer up-sampling and spatial feature restoration by using a deconvolution operation to generate a deep feature map F deep consistent with the size of the input image.

[0175] In this embodiment, S7 includes the following steps:

[0176] S71. input the deep feature map F deep into an adaptive threshold module to automatically generate a spatial dynamic threshold map T according to local statistical features.adapt (x, y), perform a binarization operation on the pixel response values, generating a preliminary segmentation result map I bin

[0177] S72. Perform a morphological reconstruction operation on the preliminary segmentation result map I bin , taking advantage of the fine linear and connectivity features of microcracks, using a structuring element S rec Perform a morphological closing operation and thinning process to generate a structure-continuous microcrack region map I recon

[0178]

[0179] wherein, are dilation and erosion operations, respectively, and Thin(·) represents a thinning operation, and the structuring element S rec is set according to the width range of the target crack;

[0180] S73. Perform a connectivity domain filtering and area constraint process on the structure-reconstructed microcrack region map I recon , eliminating isolated small regions with an area less than a set threshold A min , to form a final microcrack region segmentation map I final

[0181] In this embodiment, S8 includes the following steps:

[0182] S81. Perform a connectivity domain extraction operation on the final microcrack region segmentation map I final to obtain a set of boundary contours of all independent microcrack candidate regions and calculate a set of morphological parameters for each candidate region, respectively wherein, k is the area of the kth region, is the aspect ratio, is the average response value of the internal depth feature of the region;

[0183] S82. Construct a dynamic microcrack region classification rule based on the final fine image segmentation parameters, and set the following thresholds:

[0184] Area dynamic threshold wherein α1 is an area adjustment coefficient;

[0185] Low response intensity threshold T low = α2·T final , high threshold T high = α3·λ final ·T final , wherein α2, α3 are response adjustment coefficients, respectively;

[0186] ​​​S83. According to the performance characteristics of micro crack morphology in actual highway detection, the detection results are divided into three categories: suspected micro crack area, obvious micro crack area and non-crack area:

[0187] The non-crack area classification rule is to meet any of the following conditions: A k <A thresh , or and

[0188] The suspected micro crack area classification rule is to meet all of the following conditions: A k ≥A thresh , and

[0189] The obvious micro crack area classification rule is to meet all of the following conditions: A k ≥A thresh , and

[0190] S84. The position information of the classified crack candidate area is extracted and structured output, the boundary center point coordinates, the circumscribed rectangle boundary box and the region number, the classification label and the morphology parameter of each obvious micro crack area are extracted, and the final highway pavement micro crack detection result set is constructed.

[0191] Example 1:

[0192] On the morning of November 5, 2024, 9:30, A Provincial Transportation Survey and Design Institute cooperated with a scientific research team of a certain university to carry out micro crack inspection test on G5513 Changzhang Expressway Changde East Road section. The vehicle traffic volume of the road section has increased rapidly in the past three months, and the risk of early slight structural damage has increased. In order to provide accurate data support for preventive maintenance, the implementation team decided to apply the system of the application.

[0193] During the test, the intelligent inspection vehicle equipped with the system of the application stably drove at a speed of 20 kilometers per hour, and two FLIR black and white industrial cameras (model: FLIR BFS-U3-16S2M) installed at the bottom of the vehicle synchronously collected double-channel images. The single-channel resolution is 2448×2048, and the frame rate is 8fps. On that day, from 9:30 to 12:30, a total of 12.2 kilometers were driven, and a total of 17,568 images were collected.

[0194] The system runs in real time on the vehicle-mounted edge computing terminal. First, the image acquisition module completes the size normalization and color space conversion, unifies the image resolution to 1024×512, and converts it to a gray image. After image enhancement processing, a standardized image numbered CDX-11057 is sent to the preprocessing module.

[0195] In the image, a transverse crack with a width of only about 0.35 mm exists in the center-left position. Due to the influence of natural light reflection, the grayscale of the crack area is close to the background, making it difficult for traditional detection methods to completely extract the crack boundary. However, our system automatically enters a two-stage optimization process after noise suppression. In the first stage, 30 individuals are used to perform a coarse-grained global search, initially obtaining a crack region threshold of 0.58. Subsequently, the second stage generates a crack topology map based on the preprocessed image, guiding the search individuals to perform local refinement adjustments. Finally, the fine segmentation parameters converge to X. final ={T final =0.618,K final =3,λ final =0.92}.

[0196] The system performs segmentation and enters the feature extraction module. It uses four sets of convolutional kernels with dilation rates of 1, 2, 4, and 6 to extract multi-scale features in parallel. After channel attention enhancement and decoding upsampling, a depth feature map is generated. Subsequently, after adaptive threshold mapping and morphological closure reconstruction, the image is successfully segmented to extract the microcrack region, and the system immediately enters the detection module.

[0197] The crack was identified by the system as a significant micro-crack region, with a segmented area of ​​46 pixels², an aspect ratio of 6.2, and an average depth response value of 0.91, exceeding the high threshold of 0.84 set by the detection rules. The detection results generated the following structured data:

[0198] Image ID: CDX-11057;

[0199] Detection time: 2024-11-05 10:12:46;

[0200] Crack type: Obvious microcracks;

[0201] Location information: Image coordinates center point (512, 244);

[0202] The bounding box of the outer rectangle is (478, 237, 546, 251).

[0203] Response value: 0.91;

[0204] System response time: 1.23 seconds;

[0205] After confirming the detection result, the system automatically uploads the detection data to the cloud platform through the 5G module and sends an early warning reminder to the engineering maintenance personnel responsible for the section through the SMS interface. 10 minutes later, the maintenance personnel arrives at the designated point and uses the handheld crack detector to review the site, confirming that the crack width is 0.36mm, the length is 8cm, and it is located at the right lane edge, which is a typical temperature type micro-crack initial state. At 14:00 on the same day, the area is sprayed with a marker and included in the monthly key review plan.

[0206] In addition, the system detected 132 micro-crack candidate areas during the entire road inspection process, including 78 obvious micro-cracks and 39 suspected micro-cracks. Through random inspection verification of 50 crack targets, the maintenance personnel confirmed that the accurate recognition rate reached 92%, with 4 false positives and 1 missed detection, both of which were extremely small scale reflective interference cracks. Compared with the average time of 9.4 seconds per image for manual review, the system only needs an average of 1.6 seconds, with an 18% increase in recognition accuracy and a nearly 5.9 times increase in efficiency.

[0207] Through this real road inspection case, it is fully proved that the method of the application has excellent precision and stability in actual complex road surface environment, effectively overcomes the robustness problem of traditional methods under light, texture disturbance, has all-weather, rapid deployment, high reliability engineering application potential, and is especially suitable for micro-damage automatic inspection tasks of key highway structures such as expressway sections, tunnels, and overpasses.

[0208] The application introduces a double-stage bitter fish optimization algorithm model, introduces a double-layer mechanism of global search and local refinement search in the optimization process of image segmentation parameters, realizes a gradual optimization strategy from coarse granularity to refinement, and the optimization model not only introduces a crack sensitivity factor in the global search stage to improve the response ability to weak edges, but also fuses a crack topological structure consistency function in the local refinement stage, effectively improving the structure restoration accuracy of the micro-crack area.

[0209] The deep feature encoding network combining multi-scale hollow convolution and channel attention mechanism introduced in the micro-crack feature extraction process can more comprehensively capture crack texture information of different sizes, directions and densities. By setting a set of hole rates to construct multiple parallel hollow convolution branches, and combining the response weight distribution mechanism between channels to realize the strong expression of key crack area features, especially in road images with weak contrast and complex background texture, it performs significantly better.

[0210] The application constructs an image preprocessing flow including an image edge response and a Gaussian noise suppression combined enhancement mechanism, proposes an image fusion model based on a pixel-level nonlinear mapping function, effectively realizes the strengthening of weak edge signals and the suppression of random noise, and through the introduction of the response modulation strategy of hyperbolic tangent and exponential function, the edge contribution and noise weight are adaptively adjusted while the image is enhanced, so that the micro crack boundary is clearer and the background interference is obviously reduced.

[0211] The above merely describes a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacements or changes within the technical range disclosed by the present application according to the technical solution and the inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A highway pavement micro-crack detection system based on an improved group algorithm, characterized in that, The method comprises the following steps: An image acquisition module is used to acquire a highway pavement image and complete standardization processing, and output a standardized highway pavement image; An image preprocessing module is used to perform noise suppression and edge enhancement operations on the standardized highway pavement image, and generate a preprocessed highway pavement image; A parameter optimization module is used to perform global search and local refinement search on the preprocessed highway pavement image based on a double-stage bitterfish optimization algorithm model, and output final fine image segmentation parameters; The parameter optimization module specifically comprises a two-stage grouperfish optimization algorithm model including a global search stage and a local refinement search stage, and an initialization grouperfish optimization population , a grouperfish optimization individual represents a set of preprocessed highway pavement image segmentation parameters, including a binary threshold value for preliminary segmentation of micro-crack regions , a filter kernel size , and a feature enhancement weight parameter , wherein is the size of the grouperfish optimization population; For highway pavement micro crack area segmentation detection characteristics, based on the bitter fish optimization individual Construct fitness function ; In the global search phase of the two-stage cuckoo optimization, the current optimal cuckoo optimization individual position is calculated according to the fitness function and the average position of the cuckoo optimization population and the crack sensitivity factor is introduced ; Utilizing crack sensitivity factor updating the optimal individual position of the bitterfish The crack structure convergence threshold of the global search stage is set, and the fitness value of the current optimal kuhli optimization individual is calculated according to the fitness function; when the fitness value of the current optimal kuhli optimization individual is greater than or equal to the crack structure convergence threshold, the iteration of the global search stage is terminated, and the coarse-grained image segmentation parameter for preliminary segmentation of the micro-cracks of the highway pavement is obtained , otherwise, return to continue executing the global search process; In the local refinement search stage of the two-stage bitter fish optimization algorithm model, a local bitter fish optimization individual set is constructed wherein each local bitter fish optimization individual is added by a perturbation vector from the coarse-grained image segmentation parameters is generated; Preprocessing road surface images Constructing micro crack topology graph wherein denotes a set of crack candidate points extracted by edge detection, denotes a set of crack structure edges connected by local gradient direction and distance threshold, combining topology structure to construct crack region consistency function for evaluating the matching degree of current micro crack segmentation result and micro crack real topology structure; Constructing a composite fitness function based on topology structure to construct a crack region consistency function evaluating the local bitterfish optimization individual; Local fish optimization individual The application fish optimization update strategy is used to set the update direction of local search as the current local optimal fish optimization individual The vector difference between the current fish optimization individual position And introduce a dynamic step factor to fine control the search process, which is dynamically attenuated according to the composite fitness function of the local fish optimization individual And the preset target composite fitness function The step size gradually decreases with the improvement of individual fitness. When the termination condition is met, the fitness value of the current local optimal fish optimization individual is greater than or equal to the local crack structure convergence threshold, and a final fine image segmentation parameter is output And complete the local search of the two-stage fish optimization An image segmentation module takes the preprocessed highway pavement image and the final fine image segmentation parameters as inputs, performs preliminary micro-crack region image segmentation, and generates a preliminary micro-crack candidate region map; A feature extraction module inputs the preliminary micro-crack candidate region map into a multi-scale hollow convolution module, generates a multi-scale feature response map, and performs channel attention mechanism and feature splicing processing to form a deep feature map; A region reconstruction module performs adaptive threshold binarization and morphological reconstruction on the deep feature map, and outputs a final highway pavement micro-crack region segmentation result; A detection and output module performs candidate region detection and positioning on the highway pavement micro-crack region segmentation result, judges and filters the highway pavement micro-crack region using region contour analysis, and outputs a final highway pavement micro-crack detection result containing micro-crack position and region information.

2. A highway pavement micro-crack detection method based on an improved group algorithm, applied to the highway pavement micro-crack detection system based on the improved group algorithm in claim 1, characterized in that, The method comprises the following steps: S1. Collect a highway pavement image and perform standard processing to form a standardized highway pavement image; S2. Preprocess the standardized highway pavement image to generate a preprocessed highway pavement image; S3. Construct a double-stage bitterfish optimization algorithm model based on the preprocessed highway pavement image, initialize the optimization population in the global search stage, and evaluate the image segmentation parameters according to the fitness function to obtain coarse-grained image segmentation parameters; S3 comprises the following steps: S31. Construct a two-stage grouper optimization algorithm model including a global search stage and a local refinement search stage, initialize the grouper optimization population , grouper optimization individual represents a set of preprocessed highway pavement image segmentation parameters, including the binary threshold for preliminary segmentation of micro-crack regions , filter kernel size and feature enhancement weight parameters , wherein is the size of the grouper optimization population; S32. For the highway pavement micro-crack area segmentation detection characteristics, based on the bitter fish optimization individual Construct fitness function : wherein, is a segmentation parameter of the pre-processed highway pavement image is an inter-class variance between the micro-crack region and the background region in the pre-processed highway pavement image under the action of is a segmentation parameter of the pre-processed highway pavement image is an edge saliency of the pre-processed highway pavement image under the action of is a segmentation parameter of the pre-processed highway pavement image is a connectivity index of the micro-crack region in the pre-processed highway pavement image under the action of , , are corresponding weight coefficients, respectively. S33. In the two-stage global search phase of the Bitter Fish Optimization, the fitness function is calculated for the current best Bitter Fish Optimization individual position and the average position of the Bitter Fish Optimization population and the crack sensitivity factor is introduced: ； wherein, is the image edge saliency corresponding to the current optimal koi optimization individual, is the average value of edge saliency of all koi optimization individuals in the current koi optimization population, is the crack sensitivity adjustment coefficient; S34. Utilizing a crack sensitivity factor Update the optimal individual position of the bitter fish: wherein, , respectively denote the position of the bitterfish optimized individual at the , th iteration, and are random factors; S35. Set the crack structure convergence threshold of the global search stage, and calculate the fitness value of the current optimal KFish optimization individual according to the fitness function. When the fitness value of the current optimal KFish optimization individual is greater than or equal to the crack structure convergence threshold, terminate the iteration of the global search stage, and obtain the coarse-grained image segmentation parameters for preliminary segmentation of micro-cracks on the highway pavement , otherwise return to step S33 to continue the global search process; S4. Enter the local refinement search stage of the double-stage bitterfish optimization algorithm model using the coarse-grained image segmentation parameters, and obtain the final fine image segmentation parameters through local disturbance and regulation; S4 comprises the following steps: S41. In the local refinement search stage of the two-stage bitter fish optimization algorithm model, a local bitter fish optimization individual set is constructed wherein each local bitter fish optimization individual adding a disturbance vector by the coarse-grained image segmentation parameter generate: wherein, represents a multi-dimensional Gaussian perturbation distribution with mean 0 and covariance matrix for fine-grained sampling of the local search space; S42. Preprocessing the highway pavement image Constructing micro-crack topology graph wherein denotes the set of crack candidate points extracted by edge detection, denotes the set of crack structure edges connected by local gradient direction and distance threshold, combining with the topology structure to construct the crack region consistency function for evaluating the matching degree of the current micro-crack segmentation result and the real topology structure of micro-crack: wherein, represents a micro crack segmentation region obtained under the parameter is a direction consistency angle of the crack structure edge is an indicator function for judging whether the crack structure edge endpoints are simultaneously in the candidate region;​​ S43. Construct a crack region consistency function based on the topological structure to build a composite fitness function to evaluate the local bitterfish optimization individual: wherein, to apply a local bitterfish optimization individual to the fitness function, , is a weight coefficient for the global search phase and the local refinement search phase; S44. Local Shark Optimization Individual The application of shark optimization update strategy, set the update direction of local search for the current local optimal shark optimization individual The vector difference between the current shark optimization individual position And introduce dynamic step factor to fine control the search process, the dynamic step factor according to the composite fitness function of local shark optimization individual With the preset target composite fitness function Dynamic attenuation adjustment, step size gradually decreases with the improvement of individual fitness; S45. When the termination condition is met, the fitness value of the current locally optimal fish individual is greater than or equal to the local crack structure convergence threshold, and the final fine image segmentation parameter is output and complete the local search of the two-stage fish optimization; S5. Take the preprocessed highway pavement image and the final fine image segmentation parameters as inputs, perform image segmentation, and generate a preliminary micro-crack candidate region map; S6. Input the preliminary micro-crack candidate region map into a multi-scale hollow convolution module, generate a multi-scale feature response map, and perform channel attention mechanism and feature splicing processing to form a deep feature map; S7. Perform adaptive threshold binarization processing and morphological reconstruction on the deep feature map to eliminate residual noise and repair crack continuity, and output a final highway pavement micro-crack region segmentation result; S8. Perform candidate region detection and positioning on the highway pavement micro-crack region segmentation result, judge and filter the highway pavement micro-crack region using region contour analysis, and output a final highway pavement micro-crack detection result containing micro-crack position and region information.

3. The method according to claim 2, wherein, S1 comprises the following steps: S11. Set the highway pavement image acquisition frequency, acquire the highway pavement image through the highway pavement image acquisition device, and the highway pavement image acquisition resolution is wherein, represents the highway pavement image width, represents the highway pavement image height; S12. Perform size normalization processing on the collected original highway pavement image, and uniformly scale different size highway pavement images to a standard size wherein, represents the standard highway pavement image size, and are the uniform width and height of the standard highway pavement image, respectively. S13. Perform color space conversion operation on the size-normalized highway pavement image to convert the RGB highway pavement image into a grayscale space highway pavement image; S14. The gray-scale road surface image is subjected to road surface image enhancement processing to improve the distinguishability of micro-cracks in low-contrast regions, and an enhanced road surface image is obtained and the enhanced road surface image is taken as the standardized road surface image.

4. The method according to claim 2, wherein, The S2 comprises the following steps: S21. Based on the standardized highway pavement image , a Gaussian filter is used to perform preliminary noise suppression processing on the highway pavement image, suppress random noise generated in the environmental acquisition process, and generate a Gaussian smoothed highway pavement image ; S22. Control the filtering result of the Gaussian smoothing highway pavement image by using the gradient direction of the highway pavement image and the brightness difference between adjacent pixels to generate an edge-preserving highway pavement image; S23. Perform multi-channel filtering processing on the edge-kept highway pavement image to construct a guide filtering channel in different directions respectively wherein Each direction filtering channel enhances the micro-crack linear structure through directional response and generates a set of direction-enhanced highway pavement images ; S24. performing response fusion processing on the multi-channel directional enhanced highway pavement image set to obtain a fused highway pavement image ; S25. The edge detection processing is performed on the fused highway pavement image, and a Sobel operator is used to calculate the gradient response values of the highway pavement image in horizontal and vertical directions With , an edge response map of the highway pavement image is obtained ; S26. generating an edge response map with a Gaussian smoothing road surface image performing pixel-level fusion processing to generate a pre-processed road surface image containing noise suppression and edge enhancement features : ； wherein, is a noise modulation factor for enhancing the noise suppression effect of the Gaussian smoothed road surface image, is a decay constant for regulating the decay rate of the exponential function's influence on the Gaussian smoothed road surface image in different brightness regions, is an edge enhancement scaling factor for amplifying the contribution of the edge response map to preserve the edge details of micro-cracks in the fusion process, is an edge sensitivity adjustment factor for adjusting the non-linear mapping of the edge response values by the hyperbolic tangent function to distinguish low response and high response regions, is an edge threshold parameter for setting a reference threshold of the edge response, and only when the edge response map exceeds the edge threshold parameter, the edge information is enhanced, denote the exponential function and the hyperbolic tangent function, respectively.​​​​​ 5. The method according to claim 2, wherein, The S5 comprises the following steps: S51. Preprocessing the highway pavement image With the final fine image segmentation parameters , a preliminary image segmentation operation of the highway pavement micro-cracks is performed to generate a preliminary segmentation image : wherein, represents the pixel value at coordinates after the preliminary image segmentation, is the optimal binarization threshold in the final fine image segmentation parameter for distinguishing the micro-crack region from the non-crack region; S52. Based on the filter kernel size in the final fine image segmentation parameters to the preliminary segmentation image performing a morphological closing operation to generate a connectedness-enhanced preliminary microcrack candidate region map : wherein, and respectively denote morphological dilation and erosion operations, the size of the structuring element is determined by the final fine image segmentation parameter for smoothing out disturbances and broken areas in the micro-crack candidate region map; S53. Feature enhancement weight parameter in the final fine image segmentation parameter to the preliminary micro crack candidate region map performing a micro crack region feature enhancement operation to obtain an enhanced preliminary micro crack candidate region map : wherein, is a weight coefficient for enhancing the edge feature of the micro-crack region.

6. The method according to claim 2, wherein, The S6 comprises the following steps: S61. The enhanced preliminary micro-crack candidate region map Input a multi-scale dilated convolution submodule, and set a set of dilation rates in the multi-scale dilated convolution submodule Parallelly construct four groups of dilated convolution branches, and each group of branches adopts a corresponding dilated convolution kernel with a corresponding dilation rate Perform a dilated convolution operation on the input image to generate a dilated feature map : wherein represents a void feature response at a lower position under a void fraction of 0.

5. S62. All the hollow feature maps The input channel attention submodule assigns attention weights to the multi-scale channel features extracted by different convolution branches to obtain a channel weighted response map : wherein, denotes a global average pooling combined with max pooling operation, is an inter-channel weight learning network, denotes a feature map response weight with a dilation rate of . S63. The channel weighting response map The input feature fusion decoding submodule performs channel dimension splicing to obtain a fusion feature map , and then the fusion feature map is input into the up-sampling decoding network, and a deconvolution operation is performed for layer-by-layer up-sampling and spatial feature restoration to generate a deep feature map consistent with the size of the input image .

7. The method according to claim 2, wherein the method is characterized by, The S7 comprises the following steps: S71. The deep feature map is input into an adaptive threshold module to generate a spatial dynamic threshold map according to local statistical features , and a binarization operation is performed in combination with pixel response values to generate a preliminary segmentation result map ; S72. To the initial segmentation result map Perform morphological reconstruction operation, taking advantage of the thin linear and connectedness features of micro-cracks, using a structuring element Perform morphological closing operation and thinning process to generate a micro-crack region map with continuous structure : wherein, , are respectively an inflation and a corrosion operation, denotes a refinement operation, the size of the structuring element being set according to the width range of the target crack; S73. Micro-crack region map after structure reconfiguration Perform connected component filtering and area constraint processing to remove isolated small regions with area less than a set threshold to form the final micro-crack region segmentation map .

8. The method according to claim 7, wherein the method is characterized by, The S8 comprises the following steps: S81. Final micro crack region segmentation map Perform connected component extraction operation to obtain the boundary contour set of all independent micro crack candidate regions And calculate the morphological parameter set of each candidate region respectively Wherein, is the area of the first region, is the aspect ratio, is the average response value of the internal depth feature of the region; S82. Construct a dynamic micro-crack region classification rule in combination with the final fine image segmentation parameters, and set the following threshold values: Area dynamic threshold wherein is an area adjustment factor; low threshold of response intensity , high threshold wherein are respectively response adjustment coefficients; S83. According to the performance characteristics of the micro-crack morphology in the actual highway detection, the detection result is divided into three categories: suspected micro-crack region, obvious micro-crack region and non-crack region: The non-crack region classification rule is that any one of the following conditions is met: or and ; The suspected micro-crack region classification rule is that all of the following conditions are met: , , and ; The apparent microcrack region classification rule is to satisfy all of the following conditions: , , and ; S84. Perform position information extraction and structured output on the classified crack candidate region, extract the boundary center point coordinates, the circumscribed rectangle boundary box and the region number, the classification label and the morphology parameters of each obvious micro-crack region, and constitute a final highway pavement micro-crack detection result set.

Images