A lightweight pavement crack segmentation method and device based on a topology perception mechanism

Abstract

The present application relates to the technical fields of computer vision and road intelligence detection, in particular to a lightweight pavement crack segmentation method and device based on a topological perception mechanism, which is as follows: a pavement image is block embedded to obtain multi-scale features; route perception frequency calibration scanning and structure guiding direction refinement are sequentially performed in a multi-level encoder; selective state space propagation is performed by a deformable route constructed based on space offset and ordering offset, Laplace structure and detail redistribution are performed on the aggregated features, and low-frequency structure anchor points and space and channel gating check direction residuals; after the features of all levels are aligned, competitive reconstruction is performed by combining image-level and pixel-level scale weights, non-dominant scale information is reserved through context compensation, and a crack probability map is output. The present application can reduce the mismatch between fixed scanning routes and winding crack geometries through the complementation of deformable routes and fixed routes.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent road detection technology, and in particular to a lightweight road surface crack segmentation method and device based on topology sensing mechanism. Background Technology

[0002] Road surface cracks are an important visual indicator of road structural damage. Pixel-level crack segmentation of road surface images acquired by vehicle-mounted cameras, handheld terminals, drones, or fixed inspection equipment can provide fundamental data for assessing crack length, width, connectivity, and damage severity. Cracks in actual road surfaces are typically long, winding, branching, and have low contrast, and are easily affected by joints, ruts, scratches, stains, shadows, and grain textures. Therefore, it is necessary to simultaneously maintain long-distance topological connectivity, weak boundaries, and local details.

[0003] Existing convolutional neural networks rely on local receptive fields, typically requiring deep networks or multi-scale stacking to transmit long-distance information, making them prone to crack segmentation or endpoint loss. Self-attention-based segmentation networks can establish global relationships, but their computational and storage costs increase rapidly with the number of labels, hindering deployment on high-resolution images and resource-constrained devices. Visual state-space models employ selective scanning for long-range propagation with approximately linear complexity, providing a novel approach for lightweight crack segmentation.

[0004] However, common visual state-space models typically serialize two-dimensional features along a fixed grid path. The fixed horizontal, vertical, or bidirectional scanning path is geometrically inconsistent with irregular, curved, and bifurcated cracks, potentially causing adjacent crack points on the image plane to be separated in the sequence, resulting in serialization errors. State-space propagation can also create excessive smoothing or low-frequency responses, weakening fine cracks, weak branches, and boundaries. On the other hand, directly enhancing high-frequency information can simultaneously amplify pseudo-crack interference such as textures, stains, and markings; traditional top-down additive or splicing decoding methods struggle to select appropriate scales based on image content and pixel location, potentially blurring fine cracks or overusing shallow textures.

[0005] Therefore, there is a need for a pavement crack segmentation scheme that can simultaneously solve the problems of geometric mismatch between propagation path and crack, imbalance between structural and detail frequency allocation, insufficient reliability of directional details, and ambiguity in multi-scale reconstruction with low computational overhead. This invention proposes a lightweight pavement crack segmentation method and device based on topology sensing mechanism to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a lightweight road surface crack segmentation method and device based on a topology-aware mechanism. This invention overcomes the problems of topology mismatch and detail loss in the state-space model by adaptively scanning the route, structure, and detail frequency calibration. By combining directional refinement and competitive reconstruction, accurate crack segmentation can be achieved.

[0007] On the one hand, the technical solution of this invention to solve the technical problem is a lightweight pavement crack segmentation method based on topology sensing mechanism, which includes the following steps: S1. Obtain the road surface image to be detected, preprocess the road surface image and embed it into blocks to obtain the first scale input features; Preprocessing includes resizing, pixel normalization, and optional data augmentation; block embedding is performed by convolutional block embedding to obtain the first-scale input features; S2. Construct an encoder containing multiple coding stages. In each coding stage, execute Route-Aware Frequency-Calibrated Scan (RFCS) and Structure-Guided Directional Refinement (SGDR) sequentially, and output the current scale features. Route-aware frequency calibration scanning RFCS: Construct forward routes, reverse routes, and content-adaptive deformable routes, perform selective state-space propagation, restore the propagation results of each route to the image plane, and then fuse them to obtain multi-route aggregated features; perform Laplacian structure and detail frequency calibration on the multi-route aggregated features to obtain frequency calibration features; Structural Guided Direction Refinement SGDR: Extract low-frequency structural anchor points and multi-directional high-frequency residual features from frequency calibration features, generate spatial gating weights and channel gating weights based on low-frequency structural anchor points, use dual-gating modulation of high-frequency residual features and fuse them with frequency calibration feature residuals to output stage features; The input for the first stage of the encoding phase is the first-scale input feature; if the current encoding phase is not the last stage, the output feature of the phase is downsampled and used as the input for the next encoding phase. S3. Perform competitive scale reconstruction (CSR) on the encoded features at multiple scales. Perform channel projection and spatial alignment on the current scale features output from multiple encoding stages. Generate scale competition weights by combining image-level scale preference and pixel-level scale preference. Perform weighted reconstruction based on scale competition weights and use the context features generated from all aligned features for compensation to obtain decoded features. S4. Generate a crack probability map corresponding to the road surface image based on the decoded features. The decoded features are convolved, refined, and upsampled to the input image resolution. Crack logical values ​​are generated through a binary classification prediction head. The crack probability map is obtained through the Sigmoid function. A binary crack mask is further obtained according to a preset threshold. The segmentation network that has undergone the S1 to S4 operation process is then trained and optimized. The network parameters are iteratively updated with the goal of minimizing the loss, resulting in a lightweight road surface crack segmentation network that has been trained.

[0008] S2 is as follows: The encoder in S2 consists of three encoding stages with channel numbers of 16, 32, and 64 respectively. Each encoding stage includes a set of route-aware frequency calibration scans and structure-guided direction refinement. The state dimensions of selective state space propagation are 8, 8, and 16 respectively, and the convolution kernel sizes of the offset predictor are 7, 5, and 3 respectively. Block embedding reduces the height and width of the road image to one-quarter of their original values. Downsampling is performed between encoding stages by block merging with a stride of 2.

[0009] The construction of content-adaptive deformable routes in S2 includes: The input features in the encoding stage are subjected to layer normalization and point convolution local projection to obtain local projection features. These local projection features are then input into the offset predictor, which decouples and outputs two sets of spatial offset components and one set of sorted offset components. A bounded activation function is used to apply numerical constraints to all offset components. The spatial offset components are normalized and then superimposed onto a regular reference grid. Sampling features are obtained through bilinear interpolation. The sorted offset components are superimposed onto the raster index and sorted to obtain the access order permutation and the corresponding inverse permutation. The regular reference grid is a two-dimensional coordinate grid that perfectly matches the size of the input features and is arranged in a row-first regular pattern. Each grid point corresponds to the normalized coordinates of a spatial location on the feature map. The raster index is the set of one-dimensional integer indices corresponding to each two-dimensional grid point after the two-dimensional input features are flattened into a one-dimensional sequence according to a fixed scanning order. The sampling features are serialized by permuting according to the access order to obtain a deformable route, and the propagation result is recovered by inverse permutation.

[0010] The route-aware frequency calibration scan in S2 includes selective state-space propagation for any route. The selective state-space propagation process is as follows: Hidden state update formula: , Output marking formula: , in, Indicates route identification. , Indicates the forward route. Indicates the reverse route. This indicates a content-adaptive, deformable route; This represents the sequence step index during the traversal of the route. , This indicates the total number of serialization tags in the corresponding route; and The routes are respectively In sequence steps Input and output markers at the location, For the route In sequence steps The hidden state at that location , and The routes are respectively In sequence steps The state transition matrix, input matrix, and readout matrix at the given point. For the route Fixed jump connection parameters; at least one of the state transition matrix, input matrix, and readout matrix is ​​determined by the input marker at the corresponding position; The propagation results of the forward route, the propagation results of the reverse route after reverse recovery, and the propagation results of the deformable route after inverse permutation recovery are averaged or weighted and fused to obtain the multi-route aggregation feature.

[0011] The route-aware frequency scan in S2 includes Laplace-style structural and detail frequency calibration of multi-route aggregated features: According to frequency allocation ratio , The range of values ​​is , will have The multi-route aggregation features of each channel are divided into those with The structural branching characteristics of each channel and having The complementary branching characteristics of each channel are calculated as follows: , and ; in, This represents the rounding function; The structural branch features are structurally mixed using the first reparameterizable depthwise convolution. The structural mixing result is then subjected to average pooling with a stride of 1 to obtain the low-frequency structural response. The difference between the structural mixing result and the low-frequency structural response is then projected by point convolution to obtain the Laplacian residual. The Laplacian residual is then injected into complementary branch features after point convolution projection, and detailed blending is performed using a second reparameterizable depth convolution. The low-frequency structural response and the detailed blending result are then concatenated, point convolution projected, and normalized to obtain the frequency calibration features. In this system, the kernel size of the first reparameterizable depthwise convolution is larger than that of the second reparameterizable depthwise convolution. The first reparameterizable depthwise convolution uses a 7×7 kernel, the second reparameterizable depthwise convolution uses a 3×3 kernel, and the average pooling uses a 3×3 pooling kernel.

[0012] The refinement of the structural guidance direction in S2 includes: The frequency calibration features are group normalized to obtain normalized features. The normalized features are then subjected to average pooling, 5×5 depthwise convolution, and point convolution projection to obtain low-frequency structure anchor points. The difference between the normalized feature and its average pooling result is then calculated to obtain the high-frequency residual input. The high-frequency residual input is then processed by 3×3, 1×7 and 7×1 depth convolutions respectively, and the processing results are concatenated and dot convolution projection is performed to obtain the high-frequency residual features. Low-frequency structural anchor points are concatenated with high-frequency residual features. Spatial gating weights are predicted based on the concatenation results, and channel gating weights are predicted based on the global average pooling value of the concatenation results.

[0013] S3 is as follows: In S3, the joint image-level scale preference and pixel-level scale preference generate scale competition weights, including: Each The current scale features output from each encoding stage are projected onto... Each channel is upsampled to the spatial resolution of the first encoding stage and then concatenated to obtain aligned concatenation features; based on the aligned concatenation features, a dimension of [missing information] is generated. The image-level scale preference and dimension are ×1×1. × × The pixel-level scale preference is calculated by performing a normalized exponent operation along the scale dimension, thus broadcasting the L×1×1 image-level scale preference to the same dimension as the pixel-level scale preference in the spatial dimension. The two are then added together and the normalized exponent operation is performed again along the scale dimension to obtain the scale competition weights. and These are the height and width, respectively, based on pixel-level scale preferences. Using scale competition weights The aligned features are weighted and summed to obtain the competing features, and the aligned features are used to generate the context features. The competing features and the context features are then fused to obtain the decoding features.

[0014] S4 is as follows: The decoded features are upsampled to the original resolution of the input road surface image, and a single-channel logistic value map is obtained through a binary classification prediction head. The Sigmoid activation operation is performed on the logistic value map to obtain the crack probability map. The segmentation network is trained using road surface images with pixel-level crack annotation masks as training samples. The training samples are input into the segmentation network, and after forward propagation, a predicted crack probability map is output. The joint training loss is calculated based on the predicted crack probability map and the corresponding annotation mask. The training objective function is: , in, and These are the non-negative weights of the binary cross-entropy loss and the Dice loss, respectively. ; For binary cross-entropy loss, For Dice's loss; With the optimization objective of minimizing the joint training loss, the backpropagation algorithm is used to update all learnable parameters of the network. The AdamW optimizer is used for training, with an initial learning rate and weight decay coefficient set, and the learning rate is decayed through a multinomial learning rate decay strategy. The training samples are adjusted to a uniform resolution, and data augmentation methods such as random flipping, random cropping, and color and brightness perturbation are used to improve the generalization of the model. The training is iterated until the loss converges.

[0015] It should be noted that, without departing from the core concept of this invention, the following alternative methods can be adopted: the number of encoding stages L can be 2 or more than 3, and each stage can be set with one or more sets of RFCS-SGDR; in addition to forward and reverse grid routes, the regular routes can also include vertical, diagonal, serpentine or block routes; the sorting of adaptive routes can be approximated by differentiable sorting, segmented sorting or graph traversal; the fusion of the three routes can be achieved by averaging, static weights or input related weights.

[0016] Structure-detail decomposition can be achieved using average pooling, Gaussian smoothing, Laplacian pyramid, wavelet, or learnable low-pass operators; larger and smaller kernels can be achieved using equivalent dilated convolution, separable convolution, or deployment-time reparameterized convolution; α can be preset, selected based on the dataset, or predicted by the network based on the input image or stage features. Oriented detail branches can also incorporate diagonal, dilated, or steerable filters. Spatial and channel gating can employ Sigmoid, normalized exponentiation, or other bounded weight functions.

[0017] Scale alignment can be achieved using bilinear, nearest-neighbor, or transposed convolutional upsampling; scale scoring can be achieved using convolution, multilayer perceptron, or lightweight attention structures; context compensation can be generated from all-scale concatenated features, non-dominant scale-weighted features, or residual bypass. State space propagation can be achieved using different state spaces that satisfy input-related selective updates. Any of these alternative methods can be used to implement this invention as long as they still perform the corresponding functions in geometric path adaptation, structure-detail frequency calibration, structure-guided directional refinement, and image-level-pixel-level scale reconstruction.

[0018] On the other hand, the present invention also provides a lightweight pavement crack segmentation device based on a topology sensing mechanism, which executes a lightweight pavement crack segmentation method based on a topology sensing mechanism, including the following modules: The image acquisition and preprocessing module is used to acquire the road surface image to be detected and generate the first-scale input features; The multi-stage coding module includes a route-aware frequency calibration scanning unit and a structure-guided direction refinement unit. The route-aware frequency calibration scanning unit is used to selectively propagate state space through forward routes, reverse routes, and content-adaptive deformable routes, and to perform frequency calibration based on low-frequency structural response and Laplace residuals on the recovered and fused features. The structure-guided direction refinement unit is used to modulate high-frequency residual features in multiple directions using low-frequency structural anchors, spatial gating weights, and channel gating weights. The competitive scale reconstruction module is used to align the features output from different encoding stages, generate scale competition weights by combining image-level scale preferences and pixel-level scale preferences, perform scale weighting, and use context features for compensation. The segmentation prediction module is used to generate a crack probability map based on the decoded features output by the competitive scale reconstruction module.

[0019] The route-aware frequency calibration scanning unit includes a local projection subunit, an offset prediction subunit, a regular bidirectional route construction subunit, a deformable route construction subunit, a selective state space propagation subunit, a route recovery fusion subunit, a frequency branching subunit, a Laplace residual generation subunit, and a frequency fusion subunit. Local projection subunit: used to perform layer normalization and point convolutional local projection on the features in the input encoding stage, and output the dimensionality-reduced local projection features; Migration prediction subunit: used to predict two sets of spatial migration components and one set of sorted migration components based on local projection features, and constrains the range of migration values ​​through a bounded activation function; Regular bidirectional route construction sub-unit: used to generate two fixed raster scan routes, one forward and one reverse; Deformable route construction sub-unit: used to adjust the sampling position by combining spatial offset and the access order by combining sorting offset, to generate a content-adaptive deformable scanning route; Selective state space propagation subunit: used to perform selective state space propagation along three routes: forward, reverse, and deformable, and output the one-dimensional sequence propagation results corresponding to each route; Route recovery and fusion subunit: used to recover the propagation results of the three routes to the two-dimensional image plane, and output multi-route aggregated features after averaging or weighted fusion; Frequency branch division subunit: used to divide the multi-path aggregation features into structural branches and complementary branches along the channel dimension according to a preset frequency allocation ratio; Laplace residual generation sub-unit: used to extract the low-frequency structural response from the structural branch, calculate the difference between the structural branch and the low-frequency response to obtain the Laplace residual, and inject the residual into the complementary branch; Frequency fusion subunit: used to fuse the low-frequency structural response and the complementary branch detail features after injection residuals, and output frequency calibration features after convolution projection and normalization; The competitive scale reconstruction module includes a scale projection and alignment subunit, an image-level scoring subunit, a pixel-level scoring subunit, a competitive weighting subunit, a context compensation subunit, and a convolutional thinning subunit; Scale projection and alignment sub-unit: used to uniformly project the multi-scale features output from the multi-level coding stage to the same number of channels, and upsample them to the spatial resolution of the first coding stage, thus completing feature space alignment and stitching; Image-level scoring subunit: used to generate global image-level scale preference based on alignment stitching features, representing the optimal scale prior of the entire image; Pixel-level scoring subunit: used to generate pixel-by-pixel pixel-level scale preferences based on alignment and stitching features, adapting to scale differences of cracks at different locations; Competitive weighted subunit: used to jointly generate scale competition weights by combining image-level and pixel-level scale preferences, and perform weighted summation on multi-scale aligned features to obtain the main scale competition features; Context compensation subunit: used to generate context compensation features based on all aligned features, preserving effective information of non-dominant scales suppressed by competing weights; Convolutional thinning subunit: used to fuse competing features and context-compensated features, and outputs the final decoded features after convolutional thinning.

[0020] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a lightweight pavement crack segmentation method and device based on topology-aware mechanism. It constructs an adaptive propagation route by combining deformable sampling locations and content-related access sequences, complementing stable forward and reverse regular routes. This reduces the geometric mismatch of curved and bifurcated cracks in a fixed grid sequence, facilitating the restoration of long-range connectivity across discontinuous regions. A Laplace-style structure-details redistribution is applied to the state space features of multiple routes, utilizing larger convolutional kernels to form structural support. Changes removed from structural branches are injected into complementary branches, thereby reducing excessive low-frequency amplification rather than indiscriminately amplifying high-frequency noise. Low-frequency structural anchors and dual spatial and channel gating are used to verify isotropic and axial high-frequency residuals. It enhances endpoints, narrow branches, corners, and boundaries while suppressing crack-like interference such as grooves, stains, markings, and textures; it performs competitive reconstruction by combining image-level and pixel-level scale preferences and sets context-compensated paths, enabling the selection of appropriate semantic scales under different crack widths and background conditions, while avoiding information loss caused by strict one-size-fits-all approaches; each scanning route exhibits linear computational complexity with respect to the number of pixels, and refinement and reconstruction mainly employ depthwise convolution, pointwise convolution, and pointwise weighting; under a 512×512 input resolution, the model has 0.28M to 0.30M parameters and a computational cost of 1.03GFLOPs to 1.08GFLOPs, making it suitable for vehicle-mounted, mobile, or edge inspection equipment. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0023] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0024] Example 1 like Figure 1 As shown, a lightweight pavement crack segmentation method based on topology sensing mechanism includes the following steps: S1. Obtain the road surface image to be detected, preprocess the road surface image and embed it into blocks to obtain the first scale input features; S2. Construct an encoder that includes multiple coding stages. In each coding stage, perform route-aware frequency calibration scanning and structure-guided orientation refinement in sequence, and output the current scale features. Route-aware frequency calibration scanning RFCS: Construct forward routes, reverse routes, and content-adaptive deformable routes, perform selective state-space propagation, restore the propagation results of each route to the image plane, and then fuse them to obtain multi-route aggregated features; perform Laplacian structure and detail frequency calibration on the multi-route aggregated features to obtain frequency calibration features; Structural Guided Direction Refinement SGDR: Extract low-frequency structural anchor points and multi-directional high-frequency residual features from frequency calibration features, generate spatial gating weights and channel gating weights based on low-frequency structural anchor points, use dual-gating modulation of high-frequency residual features and fuse them with frequency calibration feature residuals to output stage features; The input for the first stage of the encoding phase is the first-scale input feature; if the current encoding phase is not the last stage, the output feature of the phase is downsampled and used as the input for the next encoding phase. S3. Perform channel projection and spatial alignment on the current scale features output from multiple encoding stages, generate scale competition weights by combining image-level scale preference and pixel-level scale preference, perform weighted reconstruction based on scale competition weights, and use the context features generated from all aligned features for compensation to obtain decoded features. S4. Generate a crack probability map corresponding to the road surface image based on the decoded features, and then train and optimize the segmentation network that has performed the S1 to S4 operation process. Iteratively update the network parameters with the goal of minimizing the loss to obtain the trained lightweight road surface crack segmentation network.

[0025] In a specific implementation, S1 is as follows: The road surface image to be detected is acquired, and the road surface image is preprocessed and segmented for embedding to obtain the first-scale input features.

[0026] Specifically, acquire the RGB road surface image to be detected. Where 3 represents the three color channels RGB. and Let I represent the height and width of the road surface image I (in pixels). Preprocessing is performed on the road surface image I, including resizing and pixel normalization. Then, the input image is downsampled by a factor of four using the convolution block embedding module Φ to obtain the first-scale input features. ,in , ,in, This represents a block embedding operation, which can be implemented through convolution with a stride of 4 or overlapping block merging. .

[0027] In a specific implementation, S2 is as follows: An encoder with multiple coding stages is constructed. In each coding stage, the route-aware frequency calibration scan (RFCS) and the structure-guided direction refinement (SGDR) are executed sequentially to output the current scale features.

[0028] The encoder comprises L stages, and the feature flow of the k-th group of RFCS and SGDR at the l-th stage can be expressed as: , ; stage output = , wherein represents the number of RFCS-SGDR groups at the l-th stage; when l < L, the input of the next stage , is block merging with a step size of 2, which is used to reduce the spatial resolution and increase the number of channels; In this embodiment, L = 3, the number of channels of the three stages is (16, 32, 64), one group of RFCS-SGDR is arranged in each stage, the state dimension of the selective state space is (8, 8, 16), and the internal expansion ratio ρ of RFCS is 2; the convolution kernel size of the offset predictor in each stage is (7, 5, 3); CSR projects each stage into d = 8 channels and then performs scale competition; the above numerical values are examples for convenient implementation, and do not constitute limitations on the number of stages, the number of channels or the state dimension.

[0029] In the specific embodiment, route-aware frequency-calibrated scanning RFCS: For the input feature of the k-th group at the l-th stage, which is abbreviated as , first perform local projection, perform layer normalization and point convolution projection on X, and divide into content features along the channel and gating feature G, the content feature obtains content feature U through 3×3 depthwise convolution and SiLU activation, and the gating feature G is activated by SiLU and used for subsequent modulation.

[0030] Construct forward route, reverse route and content-adaptive deformable route: Construction of content-adaptive deformable route: offset predictor O outputs an offset tensor according to U , is a hyperbolic tangent function, which is used as a bounded activation function to constrain the numerical range of the offset; the offset predictor O is composed of a series of convolutional layers, the output is o, the first two components of o are taken as the spatial offset Δp, which is normalized by the feature map sizes H and W and then superimposed on the regular reference grid , and the sampled feature is obtained through bilinear interpolation B ; wherein, the regular reference grid is a two-dimensional coordinate grid that perfectly matches the size of the input feature and is regularly arranged in row-major order, each grid point corresponds to the normalized coordinate of a spatial position on the feature map, the third component of o is taken as the sorting offset Δs, which is flattened into a one-dimensional vector and then superimposed on the grid index raster index It is the set of one-dimensional integer indices corresponding to each two-dimensional grid point after flattening the two-dimensional input features into a one-dimensional sequence according to a fixed scanning order (such as row priority); By sorting the sequences, we obtain the access order permutation π and its corresponding inverse permutation, thereby constructing a deformable sequence that adapts to the content. , This represents the operation of flattening two-dimensional spatial features into a one-dimensional sequence. This indicates that the sequence elements are rearranged according to the permutation π.

[0031] Regular route construction: generating forward sequences and reverse sequence , This indicates that the sequence will be reversed.

[0032] Then, selective state space propagation (SSM) is performed along three routes: forward, reverse, and deformable; for any route... The selective state-space propagation process is as follows: Hidden state update formula: , Output marking formula: , in, Indicates route identification. , Indicates the forward route. Indicates the reverse route. This indicates a content-adaptive, deformable route; This represents the sequence step index during the traversal of the route. , This indicates the total number of serialization tags in the corresponding route; and The routes are respectively In sequence steps Input and output markers at the location, For the route In sequence steps The hidden state at that location , and The routes are respectively In sequence steps The state transition matrix, input matrix, and readout matrix at the given point. To read the matrix, For the route Fixed jump connection parameters; at least one of the state transition matrix, input matrix, and readout matrix is ​​determined by the input marker at the corresponding position, thereby enabling content-adaptive selective propagation; The outputs of each route are restored and fused according to the original topology to obtain the multi-route aggregated feature: , in, This represents the operation of restoring a one-dimensional sequence to two-dimensional spatial features, and... Mutually inverse; This indicates that the original spatial order is restored by inverse permutation; , , These are the complete sequences output by the three routes; the formula represents the average fusion of the propagation results of the three routes; in an alternative, learnable weighted fusion can also be used.

[0033] Next, Laplace structure and detail frequency calibration are performed on the multi-path aggregation features, setting... have Each channel is divided into structural branches according to a ratio α. and complementary branches ,in , The structural branches are processed by 7×7 reparameterizable depthwise convolution to obtain structural hybrid features. The low-frequency structural response L is obtained by using 3×3 average pooling with a step size of 1; The difference between L and L is used to obtain the Laplace residual through point convolution projection. ;Will Post-projection injection Then, a 3×3 reparameterizable depthwise convolution is performed to obtain the blended details feature. Finally, L and The concatenation is performed using point convolution and normalization to obtain... ; , , , , , The frequency allocation ratio α is used to allocate the channel capacity of structural branches and complementary branches, and is not a frequency domain hard cutoff threshold; the RFCS output is obtained through gating and residual connections: , Reparameterizable depthwise convolutions can contain multiple branches during the training phase and are merged into a single convolutional kernel during the deployment phase.

[0034] It should be noted that the Laplace structure in this invention is not the standard Laplace transform in mathematics, nor is it equivalent to the traditional Laplace pyramid decomposition. The core idea of ​​this operation is to divide the input features into structural branches and complementary branches, extract the low-frequency structural response from the structural branches, calculate the difference between the structural branches and the low-frequency structural response to obtain the detail residual, and then inject the detail residual into the complementary branch to achieve decoupling and redistribution of the structural information and detail information of the features. This is used to separate the main structure of the crack from the detail information of the crack edge, and the direction residual is verified by low-frequency structural anchor points and spatial and channel gating.

[0035] In a specific implementation, the structural guidance direction is refined: Group normalization of the RFCS output Z yields Structural branch pairs Perform average pooling with a stride of 1, followed by 5×5 depthwise convolution and pointwise convolution projection to obtain low-frequency structural anchors. Detailed branch calculation Remove local smoothing components; Detailed branch calculation Among them, 3×3 convolutions are used to respond to corners, intersections, and endpoints, while 1×7 and 7×1 axial convolutions are used to provide complementary directional contexts.

[0036] , , , Space Gating Pixel-by-pixel selection of details compatible with structural anchor points, channel gating By controlling the level of detail in different directions and semantic channels, the residual form preserves the long-range context established by RFCS and injects only directional evidence that has passed structural verification.

[0037] In a specific implementation, S3 is as follows: Channel projection and spatial alignment are performed on the current scale features output from multiple encoding stages. Scale competition weights are generated by combining image-level scale preference and pixel-level scale preference. Weighted reconstruction is performed based on the scale competition weights, and the context features generated from all aligned features are used for compensation to obtain the decoded features.

[0038] For the output of stage l Execution scale projection get And upsampled to ( using bilinear interpolation) , ), to obtain alignment features All aligned features are stitched together along the channel dimension to obtain... ; Generate image-level scale preference g and pixel-level scale preference Q based on the alignment and stitching features U: , , in, It is an image-level scorer, consisting of a global average pooling layer and an MLP, representing the global preference of the entire image for each scale; It is a pixel-level scorer, composed of point convolutional layers, representing the preference of each spatial location for different scales; This indicates that a Softmax normalization operation is performed along the scale dimension L, such that the sum of the scale components of g is 1, and the sum of the scale components of Q at each spatial location is 1. After adding the image-level scale preference g and the pixel-level scale preference Q, perform a softmax operation along the scale dimension again to obtain the scale competition weight w: , Where g broadcasts in the spatial dimension to match the Q dimension; The competing features are obtained by weighting and summing L aligned features using scale competition weights w. : , in, This indicates element-wise multiplication. This represents the final competitive weight at each pixel location for the l-th scale; Simultaneously, context is generated from alignment and splicing features. , This represents a context compensator composed of convolutional layers, used to preserve effective information from non-dominant scales suppressed by competing weights; it also incorporates competing features. , .

[0039] In a specific implementation, S4 is as follows: Will Upsampling via bilinear interpolation The original resolution (H, W) of the input road surface image is then processed by a binary classification prediction head. (Implemented by a point convolutional layer, with 1 output channel) to obtain a single-channel logic value graph. ; Perform Sigmoid activation operation on the logic value graph S The crack probability map is obtained. Based on the preset threshold τ∈(0,1), the crack probability map can be generated. Further converted into a binary crack mask, If the value of each pixel is greater than or equal to τ, the corresponding binary crack mask is set to 1; otherwise, it is set to 0. 1 represents the real crack and 0 represents the background.

[0040] The segmentation network is trained by constructing training samples containing road surface images and corresponding binary crack masks. Images are uniformly resized to 512×512 pixels, and random flipping, cropping, and color or brightness perturbations are performed. The AdamW optimizer is used for 50 training epochs with a batch size of 1 and an initial learning rate of... The weights are decayed to 0.01, and PolyLR is used to decay the learning rate to... The random seed can be fixed at 42. No auxiliary supervision or post-processing is required during the inference phase. The loss function is calculated as follows: , , , Where N is the number of pixels. Let be the predicted crack probability for the i-th pixel. Let ε be the predicted probability of the i-th pixel, and let ε be a smoothing constant to prevent the denominator from being zero. Binary cross-entropy provides pixel-by-pixel classification constraints, and Dice loss alleviates class imbalance caused by the low proportion of crack pixels.

[0041] Example 2 A lightweight pavement crack segmentation device based on topology sensing mechanism executes a lightweight pavement crack segmentation method based on topology sensing mechanism, comprising the following modules: The image acquisition and preprocessing module is used to acquire the road surface image to be detected and generate the first-scale input features; The multi-stage coding module includes a route-aware frequency calibration scanning unit and a structure-guided direction refinement unit; The route-aware frequency calibration scanning unit is used for selective state-space propagation through forward routes, reverse routes, and content-adaptive deformable routes, and performs frequency calibration based on low-frequency structural response and Laplace residuals on the recovered fused features. The structure-guided direction refinement unit is used to modulate the high-frequency residual characteristics in multiple directions by utilizing low-frequency structural anchor points, spatial gating weights, and channel gating weights. Specifically, the route-aware frequency calibration scanning unit includes the following sub-units: Local projection subunit: used to perform layer normalization and point convolutional local projection on the features in the input encoding stage, and output the dimensionality-reduced local projection features; Migration prediction subunit: used to predict two sets of spatial migration components and one set of sorted migration components based on local projection features, and constrains the range of migration values ​​through a bounded activation function; Regular bidirectional route construction sub-unit: used to generate two fixed raster scan routes, one forward and one reverse; Deformable route construction sub-unit: used to adjust the sampling position by combining spatial offset and the access order by combining sorting offset, to generate a content-adaptive deformable scanning route; Selective state space propagation subunit: used to perform selective state space propagation along three routes: forward, reverse, and deformable, and output the one-dimensional sequence propagation results corresponding to each route; Route recovery and fusion subunit: used to recover the propagation results of the three routes to the two-dimensional image plane, and output multi-route aggregated features after averaging or weighted fusion; Frequency branch division subunit: used to divide the multi-path aggregation features into structural branches and complementary branches along the channel dimension according to a preset frequency allocation ratio; Laplace residual generation sub-unit: used to extract the low-frequency structural response from the structural branch, calculate the difference between the structural branch and the low-frequency response to obtain the Laplace residual, and inject the residual into the complementary branch; Frequency fusion subunit: used to fuse the low-frequency structural response and the complementary branch detail features after injection residuals, and output frequency calibration features after convolution projection and normalization.

[0042] The competitive scale reconstruction module is used to align the features output from different encoding stages, generate scale competition weights by combining image-level scale preferences and pixel-level scale preferences, perform scale weighting, and use context features for compensation. Specifically, the competitive scaling reconstruction module includes the following sub-units: Scale projection and alignment sub-unit: used to uniformly project the multi-scale features output from the multi-level coding stage to the same number of channels, and upsample them to the spatial resolution of the first coding stage, thus completing feature space alignment and stitching; Image-level scoring subunit: used to generate global image-level scale preference based on alignment stitching features, representing the optimal scale prior of the entire image; Pixel-level scoring subunit: used to generate pixel-by-pixel pixel-level scale preferences based on alignment and stitching features, adapting to scale differences of cracks at different locations; Competitive weighted subunit: used to jointly generate scale competition weights by combining image-level and pixel-level scale preferences, and perform weighted summation on multi-scale aligned features to obtain the main scale competition features; Context compensation subunit: used to generate context compensation features based on all aligned features, preserving effective information of non-dominant scales suppressed by competing weights; Convolutional thinning subunit: used to fuse competing features and context-compensated features, and outputs the final decoded features after convolutional thinning; The segmentation prediction module generates a crack probability map based on the decoded features output by the competitive scale reconstruction module. This module upsamples the decoded features to the input image resolution, outputs the crack probability map through a binary classification prediction head and a sigmoid function, and can further generate a binary mask.

[0043] Example 3 A computer device includes a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement a lightweight pavement crack segmentation method based on a topology-sensing mechanism.

[0044] The computer device may be an in-vehicle computing unit, an edge server, a mobile terminal, a desktop computer, or a cloud server, etc. The memory is used to store the operating system, model parameters, and computer programs; the processor is used to execute the programs to complete the method of the present invention; the device also includes an input interface and an output interface, the input interface being connected to an image acquisition device or a network, and the output interface outputting crack probability maps, masks, or inspection results.

[0045] Example 4 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a lightweight pavement crack segmentation method based on a topology-sensing mechanism.

[0046] The computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), flash memory, solid-state drive (SSD), magnetic disk, optical disk, or other non-transitory medium capable of storing program instructions. When the program instructions are executed by one or more processors, they cause the processors to perform the method of any of the above embodiments.

[0047] Example 5 This embodiment further demonstrates the beneficial effects of the technical solution of the present invention through experimental data.

[0048] The proposed method was validated on four publicly available pavement crack datasets: Crack500, DeepCrack, CrackMap, and CamCrack789, with an input resolution of 512×512 for each dataset. Evaluation metrics included precision (P), recall (R), F1 score, best-in-class F1 score (ODS) for the dataset, mean of best-in-class F1 scores (OIS) for images, and mean intersection-over-union ratio (mIoU). Different frequency allocation ratios α were used for different datasets: 0.2 for Crack500, 0.1 for DeepCrack and CrackMap, and 0.4 for CamCrack789. The model parameters ranged from 0.28M to 0.30M, and the computational cost ranged from 1.03 GFLOPs to 1.08 GFLOPs.

[0049] Table 1 shows the segmentation results and computational costs of this invention on four datasets.

[0050] Table 1. Segmentation results and computational costs for the four datasets. The results show that the present invention maintains high structural integrity and segmentation accuracy on all four datasets, with extremely low computational overhead.

[0051] Table 2 shows the ablation experiment results of the main modules, verifying the effectiveness of each core module.

[0052] Table 2 Ablation results of the main modules Ablation experiments show that RFCS primarily improves structural continuity on irregular paths; SGDR utilizes structural anchors to reduce the over-response of directional details to background interference; and CSR further eliminates scale ambiguity through image-to-pixel scale competition and context compensation. The sequential addition of each module leads to performance improvements, demonstrating the synergistic effect and effectiveness of each step in the technical solution of this invention.

[0053] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A lightweight pavement crack segmentation method based on topology sensing mechanism, characterized in that, Includes the following steps: S1. Obtain the road surface image to be detected, preprocess the road surface image and embed it into blocks to obtain the first scale input features; S2. Construct an encoder that includes multiple coding stages. In each coding stage, perform route-aware frequency calibration scanning and structure-guided orientation refinement in sequence, and output the current scale features. Route-aware frequency calibration scanning: Construct forward routes, reverse routes, and content-adaptive deformable routes, carry out selective state-space propagation, and then fuse the propagation results of each route back to the image plane to obtain multi-route aggregated features. Laplace structure and detail frequency calibration are performed on the multi-path aggregation features to obtain frequency calibration features; Structural guidance direction refinement: Low-frequency structural anchor points and multi-directional high-frequency residual features are extracted from frequency calibration features. Spatial gating weights and channel gating weights are generated based on low-frequency structural anchor points. High-frequency residual features are modulated using dual gating and fused with frequency calibration feature residuals to output stage features. The input for the first stage of the encoding phase is the first-scale input feature; If the current encoding stage is not the last stage, the output features of the stage are downsampled and used as the input for the next encoding stage. S3. Perform channel projection and spatial alignment on the current scale features output from multiple encoding stages, generate scale competition weights by combining image-level scale preference and pixel-level scale preference, perform weighted reconstruction based on scale competition weights, and use the context features generated from all aligned features for compensation to obtain decoded features. S4. Generate a crack probability map corresponding to the road surface image based on the decoded features, and then train and optimize the segmentation network that has performed the S1 to S4 operation process. Iteratively update the network parameters with the goal of minimizing the loss to obtain the trained lightweight road surface crack segmentation network.

2. The lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, The encoder in S2 includes three coding stages with channel numbers of 16, 32 and 64 respectively. Each coding stage includes a set of route-aware frequency calibration scans and structure-guided orientation refinement. The state dimensions of the selective state space propagation are 8, 8 and 16 respectively, and the convolution kernel sizes of the offset predictor are 7, 5 and 3 respectively. Block embedding reduces the height and width of the road image to one-quarter of their original values, and downsampling is performed between each coding stage by merging blocks with a step size of 2.

3. The lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, The construction of content-adaptive deformable routes in S2 includes: The input features in the encoding stage are subjected to layer normalization and point convolution local projection to obtain local projection features. These local projection features are then input into the offset predictor, which decouples and outputs two sets of spatial offset components and one set of sorted offset components. The tanh bounded activation function is used to numerically constrain all offset components, limiting them to the interval [-1, 1]. The spatial offset components are normalized and then superimposed onto a regular reference grid. Sampling features are obtained through bilinear interpolation. The sorted offset components are superimposed onto the raster index and sorted to obtain the access order permutation and the corresponding inverse permutation. The regular reference grid is a two-dimensional coordinate grid that perfectly matches the size of the input features and is arranged in a row-first regular pattern. Each grid point corresponds to the normalized coordinates of a spatial location on the feature map. The raster index is the set of one-dimensional integer indices corresponding to each two-dimensional grid point after the two-dimensional input features are flattened into a one-dimensional sequence according to a fixed scanning order. The sampling features are serialized by permuting according to the access order to obtain a deformable route, and the propagation result is recovered by inverse permutation.

4. The lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, The route-aware frequency calibration scan in S2 includes selective state-space propagation for any route. The selective state-space propagation process is as follows: Hidden state update formula: ， Output marking formula: ， in, Indicates route identification. , Indicates the forward route. Indicates the reverse route. This indicates a content-adaptive, deformable route; This represents the sequence step index during the traversal of the route. , This indicates the total number of serialization tags in the corresponding route; and The routes are respectively In sequence steps Input and output markers at the location, For the route In sequence steps The hidden state at that location , and The routes are respectively In sequence steps The state transition matrix, input matrix, and readout matrix at the given point. For the route Fixed jump connection parameters; at least one of the state transition matrix, input matrix, and readout matrix is ​​determined by the input marker at the corresponding position; The propagation results of the forward route, the propagation results of the reverse route after reverse recovery, and the propagation results of the deformable route after inverse permutation recovery are averaged or weighted and fused to obtain the multi-route aggregation feature.

5. A lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, The route-aware frequency scan in S2 includes Laplace-style structural and detail frequency calibration of multi-route aggregated features: According to frequency allocation ratio , The range of values ​​is , will have The multi-route aggregation features of each channel are divided into those with The structural branching characteristics of each channel and having The complementary branching characteristics of each channel are calculated as follows: ， and ; in, This represents the rounding function; The structural branch features are structurally mixed using the first reparameterizable depthwise convolution. The structural mixing result is then subjected to average pooling with a stride of 1 to obtain the low-frequency structural response. The difference between the structural mixing result and the low-frequency structural response is then projected by point convolution to obtain the Laplacian residual. The Laplacian residual is then injected into complementary branch features after point convolution projection, and detailed blending is performed using a second reparameterizable depth convolution. The low-frequency structural response and the detailed blending result are then concatenated, point convolution projected, and normalized to obtain the frequency calibration features. In this system, the kernel size of the first reparameterizable depthwise convolution is larger than that of the second reparameterizable depthwise convolution. The first reparameterizable depthwise convolution uses a 7×7 kernel, the second reparameterizable depthwise convolution uses a 3×3 kernel, and the average pooling uses a 3×3 pooling kernel.

6. A lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, The refinement of the structural guidance direction in S2 includes: The frequency calibration features are group normalized to obtain normalized features. The normalized features are then subjected to average pooling, 5×5 depthwise convolution, and point convolution projection to obtain low-frequency structure anchor points. The difference between the normalized feature and its average pooling result is then calculated to obtain the high-frequency residual input. The high-frequency residual input is then processed by 3×3, 1×7 and 7×1 depth convolutions respectively, and the processing results are concatenated and dot convolution projection is performed to obtain the high-frequency residual features. Low-frequency structural anchor points are concatenated with high-frequency residual features. Spatial gating weights are predicted based on the concatenation results, and channel gating weights are predicted based on the global average pooling value of the concatenation results.

7. A lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, In S3, the joint image-level scale preference and pixel-level scale preference generate scale competition weights, including: Each The current scale features output from each encoding stage are projected onto... Each channel is upsampled to the spatial resolution of the first encoding stage and then concatenated to obtain aligned concatenation features; based on the aligned concatenation features, a dimension of [missing information] is generated. The image-level scale preference and dimension are ×1×1. × × The pixel-level scale preference is calculated by performing a normalized exponent operation along the scale dimension, thus broadcasting the L×1×1 image-level scale preference to the same dimension as the pixel-level scale preference in the spatial dimension. The two are then added together and the normalized exponent operation is performed again along the scale dimension to obtain the scale competition weights. and These are the height and width, respectively, based on pixel-level scale preferences. Using scale competition weights The aligned features are weighted and summed to obtain the competing features, and the aligned features are used to generate the context features. The competing features and the context features are then fused to obtain the decoding features.

8. A lightweight pavement crack segmentation method based on topology sensing mechanism according to claim 1, characterized in that, S4 is as follows: The decoded features are upsampled to the original resolution of the input road surface image, and a single-channel logistic value map is obtained through a binary classification prediction head. The Sigmoid activation operation is performed on the logistic value map to obtain the crack probability map. The segmentation network is trained using road surface images with pixel-level crack annotation masks as training samples. The training samples are input into the segmentation network, and after forward propagation, a predicted crack probability map is output. The joint training loss is calculated based on the predicted crack probability map and the corresponding annotation mask. The training objective function is: ， in, and These are the non-negative weights of the binary cross-entropy loss and the Dice loss, respectively. ; For binary cross-entropy loss, For Dice's loss; With the optimization objective of minimizing the joint training loss, the backpropagation algorithm is used to update all learnable parameters of the network. The AdamW optimizer is used for training, with an initial learning rate and weight decay coefficient set, and the learning rate is decayed through a multinomial learning rate decay strategy. The training samples are adjusted to a uniform resolution, and data augmentation methods such as random flipping, random cropping, and color and brightness perturbation are used to improve the generalization of the model. The training is iterated until the loss converges.

9. A lightweight pavement crack segmentation device based on topology sensing mechanism, performing the lightweight pavement crack segmentation method based on topology sensing mechanism as described in any one of claims 1-8, characterized in that, Includes the following modules: The image acquisition and preprocessing module is used to acquire the road surface image to be detected and generate the first-scale input features; The multi-stage coding module includes a route-aware frequency calibration scanning unit and a structure-guided direction refinement unit. The route-aware frequency calibration scanning unit is used to selectively propagate state space through forward routes, reverse routes, and content-adaptive deformable routes, and to perform frequency calibration based on low-frequency structural response and Laplace residuals on the recovered and fused features. The structure-guided direction refinement unit is used to modulate high-frequency residual features in multiple directions using low-frequency structural anchors, spatial gating weights, and channel gating weights. The competitive scale reconstruction module is used to align the features output from different encoding stages, generate scale competition weights by combining image-level scale preferences and pixel-level scale preferences, perform scale weighting, and use context features for compensation. The segmentation prediction module is used to generate a crack probability map based on the decoded features output by the competitive scale reconstruction module.

10. A lightweight pavement crack segmentation device based on topology sensing mechanism according to claim 9, characterized in that: The route-aware frequency calibration scanning unit includes a local projection subunit, an offset prediction subunit, a regular bidirectional route construction subunit, a deformable route construction subunit, a selective state space propagation subunit, a route recovery fusion subunit, a frequency branching subunit, a Laplace residual generation subunit, and a frequency fusion subunit. Local projection subunit: used to perform layer normalization and point convolutional local projection on the features in the input encoding stage, and output the dimensionality-reduced local projection features; Migration prediction subunit: used to predict two sets of spatial migration components and one set of sorted migration components based on local projection features, and constrains the range of migration values ​​through a bounded activation function; Regular bidirectional route construction sub-unit: used to generate two fixed raster scan routes, one forward and one reverse; Deformable route construction sub-unit: used to adjust the sampling position by combining spatial offset and the access order by combining sorting offset, to generate a content-adaptive deformable scanning route; Selective state space propagation subunit: used to perform selective state space propagation along three routes: forward, reverse, and deformable, and output the one-dimensional sequence propagation results corresponding to each route; Route recovery and fusion subunit: used to recover the propagation results of the three routes to the two-dimensional image plane, and output multi-route aggregated features after averaging or weighted fusion; Frequency branch division subunit: used to divide the multi-path aggregation features into structural branches and complementary branches along the channel dimension according to a preset frequency allocation ratio; Laplace residual generation sub-unit: used to extract the low-frequency structural response from the structural branch, calculate the difference between the structural branch and the low-frequency response to obtain the Laplace residual, and inject the residual into the complementary branch; Frequency fusion subunit: used to fuse the low-frequency structural response and the complementary branch detail features after injection residuals, and output frequency calibration features after convolution projection and normalization; The competitive scale reconstruction module includes a scale projection and alignment subunit, an image-level scoring subunit, a pixel-level scoring subunit, a competitive weighting subunit, a context compensation subunit, and a convolutional thinning subunit; Scale projection and alignment sub-unit: used to uniformly project the multi-scale features output from the multi-level coding stage to the same number of channels, and upsample them to the spatial resolution of the first coding stage, thus completing feature space alignment and stitching; Image-level scoring subunit: used to generate global image-level scale preference based on alignment stitching features, representing the optimal scale prior of the entire image; Pixel-level scoring subunit: used to generate pixel-by-pixel pixel-level scale preferences based on alignment and stitching features, adapting to scale differences of cracks at different locations; Competitive weighted subunit: used to jointly generate scale competition weights by combining image-level and pixel-level scale preferences, and perform weighted summation on multi-scale aligned features to obtain the main scale competition features; Context compensation subunit: used to generate context compensation features based on all aligned features, preserving effective information of non-dominant scales suppressed by competing weights; Convolutional thinning subunit: used to fuse competing features and context-compensated features, and outputs the final decoded features after convolutional thinning.

Images