A pavement crack detection method based on edge perception and global receptive field modeling

Abstract

This invention relates to the fields of computer vision and road infrastructure inspection, specifically to a method for detecting road surface cracks based on edge perception and global receptive field modeling. The steps include: inputting an image into an encoder-decoder network for multi-scale feature extraction; inputting feature maps of at least one level into a biomimetic edge perception module for edge feature enhancement; inputting feature maps of other levels into a global receptive field module for enhancement processing; and upsampling the enhanced features level by level through the decoder, fusing the previous level's decoded features with the enhanced features output by the global receptive field module to generate a road crack segmentation map. By introducing the concept of biomimetic visual edge perception, the model's ability to perceive road cracks is enhanced. Combined with a global receptive field modeling strategy, it achieves joint modeling of the relationship between local details and the overall spatial structure of cracks, thereby improving the detection accuracy and stability under complex road surface background conditions.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and road infrastructure inspection, and in particular to a method for detecting road surface cracks based on edge perception and global receptive field modeling. Background Technology

[0002] Modern highways are mostly constructed of asphalt concrete or cement concrete. Over time and with repeated traffic loads, various types of cracks inevitably appear on the road surface, accelerating the deterioration of the road structure. Timely and accurate detection of highway cracks is crucial in highway maintenance. Currently, most highway crack detection still relies on manual inspections or vehicle-mounted inspection equipment. Manual inspections suffer from low efficiency, high missed detection rates, high operational risks, and are greatly influenced by the subjective experience of the inspectors. While vehicle-mounted inspections offer higher accuracy, their accessibility in remote sections or complex terrain is limited, and the equipment is expensive.

[0003] In recent years, drones, equipped with high-definition imaging equipment, have acquired large-scale, high-resolution highway image data, providing new technical means and application scenarios for automated highway crack detection. While deep learning, especially convolutional neural networks (CNNs), has made significant progress in image recognition and object detection with the rapid development of artificial intelligence, it still faces serious challenges in highway crack detection applications.

[0004] Crack detection in highway imagery acquired by UAVs faces two major challenges: First, the target characteristics are complex. In low-altitude aerial imagery, cracks (especially micro-cracks) typically appear as elongated structures only a few pixels wide, with highly similar and weakly discriminative shallow features, resulting in low detection sensitivity. Cracks also show only slight differences in color and texture from the road surface background, especially at a vertical overhead view, making visual differentiation extremely low. Second, environmental interference is significant. Highway backgrounds often contain various interfering elements such as road markings, joints, oil stains, and shadows, which can easily be confused with real cracks. Therefore, improving the detection accuracy and robustness of small cracks in UAV highway imagery under complex backgrounds and variable imaging conditions remains a critical issue that current technologies urgently need to address.

[0005] Related research indicates that biological vision systems possess unique advantages in perceiving faint targets and complex environments. For example, ganglion cells in the frog's retina exhibit a persistent response to abrupt changes in brightness or contrast at the edges of its visual field. This visual principle, known as "Sustained Contrast Detection (SCD)," enables it to efficiently capture small, low-contrast targets in cluttered environments. This biological vision mechanism bears a striking resemblance to crack detection in UAV road imagery. Therefore, drawing inspiration from the edge and contrast perception mechanisms in biological vision provides valuable insights for improving crack detection performance.

[0006] However, existing biomimetic detection methods mostly remain at the module splicing level, and fail to truly achieve a deep integration of biological vision mechanisms and deep network architecture by simply introducing edge operators or attention mechanisms. In specific applications, they still face three major technical bottlenecks: (1) Crack targets are extremely small in scale (often only a few pixels wide) and have a highly continuous structure, making it difficult for traditional local convolution or single edge detection to stably capture their complete topology; (2) Biomimetic edge perception is usually limited to the local receptive field and lacks the ability to model long-distance context, resulting in insufficient understanding of the overall structure of extended cracks; (3) Under complex road surface interference (such as markings, oil stains, shadows) and multi-variable imaging conditions (lighting, scale, blur), it is easy to produce false detections or missed detections, making it difficult to balance high recall and low false alarm rate. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems of low crack detection accuracy, poor robustness, and insufficient accuracy and stability in existing technologies. It proposes for the first time a unified structural and synergistic biomimetic-global fusion architecture. By introducing a biomimetic visual mechanism and a multi-scale global modeling strategy, it provides a road surface crack detection method based on edge perception and global receptive field modeling. This method employs an encoder-decoder deep neural network structure, which, without significantly increasing computational complexity, simultaneously improves sensitivity to crack edge details and the ability to understand the overall road scene structure. The steps include:

[0008] S1. Input the image to be detected into the encoder-decoder network, and extract multi-scale features from the input image through the encoder to obtain feature maps at different levels;

[0009] S2. Input the feature map of at least one of the layers in the encoder into the bionic edge perception module to enhance the edge features; input the feature map of at least one other layer into the corresponding global receptive field module to perform context aggregation and detail enhancement fusion processing, and output the enhanced features;

[0010] S3. The edge features and the enhancement features are upsampled step by step by the decoder to reconstruct the image spatial resolution, and at at least one decoding level, the decoding features of the previous level are fused with the enhancement features output by the global receptive field module of the corresponding level; based on the fused decoding features of all levels, a multi-scale highway crack segmentation map is generated.

[0011] Preferably, the biomimetic edge perception module in step S2 includes dual-path contrast enhancement and spatial attention weight generation.

[0012] Preferably, the dual-path contrast enhancement is used to enhance the feature map, and the process includes:

[0013] The feature map is extracted in parallel using two-path features, wherein the first path extracts the edge features and the second path extracts global statistical features;

[0014] The edge features and the global statistical features are concatenated along the channel dimension to obtain the fused features;

[0015] After channel expansion, the fused features are added element-wise to the original input feature map in the spatial domain to output the enhanced edge features.

[0016] Preferably, the first path extracts complementary edge features by applying the Sobel and Canny operators in parallel to the feature map; the second path extracts global statistical features from the feature map by global average pooling and global max pooling.

[0017] Preferably, the spatial attention weight generation takes the output of the dual-path contrast enhancement as input, generates an initial attention map through a lightweight convolutional layer and a nonlinear activation function, and then normalizes it to obtain a spatial weight map.

[0018] Preferably, the global receptive field module in step S2 includes wide-domain context aggregation and local detail enhancement. After processing the feature map, it is concatenated in the channel dimension and fused and output through 1×1 convolution to generate the enhanced feature.

[0019] The wide-domain context aggregation includes dilated convolution parallel branches to capture multi-scale contextual information;

[0020] The local detail enhancement is achieved through lightweight convolution operations to preserve the local detail information of the feature map.

[0021] Preferably, the wide-domain context aggregation includes at least four parallel branches, each of which employs dilated convolutions with different dilation rates, the set of dilation rates being {1, 2, 4, 8}.

[0022] Preferably, the wide-domain context aggregation further uses a channel attention mechanism to adaptively generate fusion weights for the output features of each of the parallel branches.

[0023] Preferably, the local detail enhancement process includes:

[0024] Dimensionality reduction is performed on the feature map;

[0025] Perform continuous depthwise separable convolutions on the reduced feature map;

[0026] The result of the depthwise separable convolution is added to the feature map of the original input via residual connections.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention provides a method for detecting road surface cracks. The biomimetic edge perception module, designed using biomimetic principles, improves the sensitivity to high-frequency edge structures of small-scale cracks by simulating the continuous contrast detection visual mechanism of a frog, while effectively suppressing background interference and improving detection accuracy.

[0029] 2. The road surface crack detection method provided by this invention also overcomes the problem of limited receptive field in traditional local convolution by using a feature fusion module based on global receptive field modeling. This enables joint modeling of the spatial distribution of cracks and the overall structural relationship of the road, thereby effectively suppressing interference from complex backgrounds and improving the accuracy and robustness of UAV road crack detection. Attached Figure Description

[0030] Figure 1 This is a general principle framework diagram of an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the biomimetic edge sensing module structure according to an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the global receptive field modeling module structure in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram illustrating the identification of small, low-contrast road cracks in an embodiment of the present invention.

[0034] Figure 5 This is a schematic diagram illustrating the identification of highway cracks in an urban road background, as described in an embodiment of the present invention.

[0035] Figure 6 This is a schematic diagram illustrating the identification of highway cracks under harsh conditions in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0037] Unless otherwise specified, the use of terms such as "upper," "lower," "left," "right," "center," "inner," and "outer" to indicate orientation or positional relationships in the description of specific embodiments of the present invention is based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationship in which the product / equipment / device is typically placed during use. These terms are merely for the purpose of facilitating the description of the present invention or simplifying the description in specific embodiments, enabling those skilled in the art to quickly understand the solution, and do not indicate or imply that a particular device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship. Therefore, they should not be construed as limitations on the present invention.

[0038] Furthermore, the use of terms such as "horizontal," "vertical," "suspended," and "parallel" does not imply that the corresponding device / component / element must be absolutely horizontal, vertical, suspended, or parallel, but rather that it can be slightly tilted or have a deviation. For example, "horizontal" merely means that its direction is more horizontal relative to "vertical," not that the structure must be completely horizontal, but that it can be slightly tilted. Alternatively, it can be simplified to mean that the corresponding device / component / element, when set in a "horizontal," "vertical," "suspended," or "parallel" direction, can have an error / deviation of ±10% relative to the corresponding direction, more preferably within ±8%, more preferably within ±6%, more preferably within ±5%, and more preferably within ±4%. As long as the corresponding device / component / element is within the error / deviation range, it can still achieve its function in the present invention.

[0039] Furthermore, the use of terms such as "first," "second," "third," etc. in terminology is merely for distinguishing identical or similar components and should not be interpreted as emphasizing or implying the relative importance of a particular component.

[0040] Furthermore, in the description of the embodiments of the present invention, "several", "more than", and "a number of" represent at least two. The number can be any number, such as 2, 3, 4, 5, 6, 7, 8, or 9, and can even exceed nine.

[0041] Furthermore, in the description of the technical solution of this invention, unless otherwise explicitly specified / limited / restricted, the terms "set up," "install," "connect," "link," "provided with," "laid out," and "arranged" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to common connection methods in the art, such as welding, riveting, bolting, and threaded connections. Such connections can be mechanical, electrical, or communication connections; they can be direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components.

[0042] Example 1

[0043] This embodiment provides a road surface crack detection method based on edge perception and global receptive field modeling. Its overall implementation principle framework is as follows: Figure 1 As shown, the implementation is as follows: The input image (width, height) first enters the encoder section, passing sequentially through encoders 1 to 4 (resolutions of width / 4, height / 4; width / 8, height / 8; width / 16, height / 16; width / 32, height / 32, respectively), extracting multi-scale features step by step; the output of encoder 4 is simultaneously fed into the "bionic perception module" to enhance high-frequency edge responses; while the outputs of encoders 2, 3, and 4 are respectively connected to the "global receptive field module," which includes a "wide-domain context aggregation module" and a "local detail enhancement module." The former models long-range semantic dependencies through multi-branch dilated convolution. The latter preserves fine-grained structure through depthwise separable convolution, and the fusion of the two outputs enhanced features. Subsequently, the features of each level of the encoder are upsampled stepwise by decoders 1 to 4 to restore spatial resolution. The inputs of decoders 2, 3, and 4, in addition to coming from the previous level decoder, also incorporate the output of the global receptive field module of the corresponding level, realizing multi-scale context guidance. Finally, all decoder outputs are processed by "convolution + sigmoid + upsampling" (i.e., the "prediction layer") to generate four crack segmentation maps P1–P4 at different scales, completing multi-scale crack detection and boundary refinement from coarse to fine. This architecture enhances edge sensitivity through a biomimetic edge perception module, overcomes local limitations through a global receptive field module, and unifies the output of the prediction layer. The three work together to achieve high-precision and robust detection of tiny cracks in complex road surface backgrounds.

[0044] Based on the above overall design and implementation ideas, the specific implementation steps and contents of this embodiment include:

[0045] S1. Input the image to be detected into the encoder-decoder network, and extract multi-scale features from the input image through the encoder to obtain feature maps at different levels;

[0046] S2. Input the feature map of at least one of the layers in the encoder into the bionic edge perception module to enhance the edge features; input the feature map of at least one other layer into the corresponding global receptive field module to perform context aggregation and detail enhancement fusion processing, and output the enhanced features;

[0047] To address the challenges of small-scale cracks, weak edge contrast, and susceptibility to complex background interference in UAV highway imagery, this embodiment incorporates a biomimetic edge perception module during the encoding stage. By mimicking the continuous contrast detection mechanism in frog retinal ganglion cells, the network's response to high-frequency edge structures of highway cracks is significantly enhanced. The overall structure of this module is as follows: Figure 2As shown, it includes a dual-path contrast enhancement module and a spatial attention weight generation module, which can effectively extract the edge features of cracks and suppress complex background interference, thereby achieving robust enhancement of crack features.

[0048] S21. Implementation of the dual-path contrast enhancement module

[0049] In UAV images of highway cracks, cracks typically appear as long, continuous edge structures, with their salient features mainly manifested in abrupt changes in local brightness or texture. To enhance the network's adaptability to different types of crack edges, a dual-path contrast enhancement module was constructed. This module extracts different forms of edge response information in parallel, achieving complementary modeling of crack edge features.

[0050] Preferably, the enhancement process of the dual-path contrast enhancement module includes:

[0051] The feature map is extracted in parallel using two paths. The first path extracts the edge features by applying the Sobel and Canny operators in parallel to extract complementary edge features. The second path extracts global statistical features from the feature map using global average pooling and global max pooling.

[0052] The edge features and the global statistical features are concatenated along the channel dimension to obtain the fused features;

[0053] After channel expansion, the fused features are added element-wise to the original input feature map in the spatial domain to output the enhanced edge features.

[0054] Specifically, for a given input feature map The Sobel and Canny operators are applied in parallel to extract complementary edge features: the Sobel operator has a good noise suppression effect and can stably extract the coarse outline and gradual gradient information of the crack; the Canny operator, through nonmaximum suppression and double threshold detection, can accurately detect the fine edges and subtle branches of the crack. Finally, the two are fused by weighted averaging and difference operations to obtain the final result. E edge .

[0055]

[0056] Among them, E fused E represents the fused edge feature map, Wq represents the spatial attention weight map, and each element takes a value in the range [0,1], which is used to dynamically adjust the fusion ratio of Sobel and Canny edges at each spatial location. S E represents the edge response map obtained after processing by the Sobel operator. CThe diagram represents the edge response map after processing by the Canny operator, and ⊙ represents the element-wise multiplication operation. This mechanism allows the model to dynamically allocate the contributions of Sobel edges and Canny edges at different spatial locations (preferring the smooth response of Sobel edges in weak edge regions), generating high-quality structural priors for subsequent spatial attention modules.

[0057] Meanwhile, another branch extracts complementary global statistical information from the input feature map through Global Average Pooling (GAP) and Global Maximum Pooling (GMP), respectively. GAP calculates the spatial mean of each channel, capturing the low-frequency characteristics of cracks in the overall distribution; GMP retains the maximum activation value of each channel, emphasizing the most salient regions and preserving strong response cues, making it particularly suitable for highlighting the high-response regions of cracks.

[0058]

[0059]

[0060] Among them, F gap F represents the global statistical feature vector obtained after global average pooling. gmp X represents the global statistical feature vector obtained after global max pooling, where X is the input feature map, and H and W are its height and width, respectively. c,i,j The max represents the pixel value at spatial location (i,j) for each of the c channels of the input feature map. i,j X c,i,j This indicates the operation of taking the maximum value for all spatial positions of the c-th channel.

[0061] The resulting edge enhancement features With global features , By concatenating along the channel dimension, a unified feature vector is generated that integrates local edge structure and global statistical context:

[0062]

[0063] in, This represents the fused global feature vector. This represents the input feature map, whose spatial dimensions are... The number of channels is , Representation of feature map In the individual channels, spatial locations Pixel value at that location, Indicates the first The average value is calculated for all spatial locations of each channel (i.e., global average pooling). Indicates the first The maximum value is taken from all spatial locations of each channel (i.e., global max pooling). This indicates a concatenation operation along the channel dimension, resulting in a final output feature vector with dimension . .

[0064] Subsequently, the splicing features are expanded in the channel dimension and combined with the original features in the spatial domain. Element-wise addition is used to enhance both local structural details and global semantic context, as shown in the following equation:

[0065]

[0066] in, This represents the output feature map after double enhancement. For the input feature map, This represents a multilayer perceptron used for fusing edge features. Perform a nonlinear transformation to match the channel dimension. This indicates that the global feature vector Reshape to the input feature map Same space size The feature map is used to perform element-wise addition operations.

[0067] S22. Implementation of the Spatial Attention Weight Generation Module

[0068] The spatial attention weight generation module uses the output of the dual-path contrast enhancement module. For input ( This module constructs a spatial attention mechanism to further focus on the crack region and suppress background interference. It first generates an initial attention map using lightweight convolutional layers and a non-linear activation function, then normalizes it using the sigmoid function to obtain a spatial weight map.

[0069]

[0070] in, and 1×1 and 3×3 represent 1×1 and 3×3 convolution operations, respectively, used to extract inter-channel dependencies and local spatial context; ReLU is the activation function; the Sigmoid function normalizes the output to the [0,1] interval, generating the spatial attention weight distribution.

[0071] This weighted map amplifies the feature response of the crack region through element-wise multiplication, while suppressing background interference regions.

[0072]

[0073] in, This represents the feature map modulated by the spatial attention weight map. This represents the fused features output by the dual-path contrast enhancement module (i.e., the features enhanced by both the aforementioned local edge structure and global statistical context). This represents the spatial attention weight map output by the spatial attention weight generation module, where each element takes a value in the range [0,1], and is used to adjust the spatial attention weight at each spatial location. Perform adaptive weighting. This indicates an element-wise multiplication operation.

[0074] To preserve important information from the original features and avoid the loss of crack details due to over-smoothing, the SAWB module introduces a residual connection mechanism:

[0075]

[0076] in, This represents the final output feature map of the spatial attention weight generation module. Features after attention weight modulation The original enhancement features input to this module are coupled element-wise to achieve residual connection, ensuring that high-frequency detail information such as crack edges is preserved.

[0077] Final output While effectively suppressing interference from complex road surface backgrounds, the design significantly enhances the characterization ability of crack edge features. This design fully simulates the continuous response characteristics of biological visual systems to brightness abrupt changes at edges, enabling the network to accurately capture the high-frequency edge structure of highway cracks and maintain high detection robustness even under harsh conditions such as changing illumination and complex backgrounds.

[0078] To overcome the inherent limitations of local receptive fields in biomimetic edge perception mechanisms, which hinder the effective capture of large-scale contextual information or the perception of distant crack structures, this embodiment proposes a crack contextual information fusion method based on global receptive field modeling. This module expands the effective receptive field through multi-scale dilated convolution and a local-global feature fusion strategy, thereby simulating an enhanced biological vision system. While maintaining a keen response to local crack details, it improves the perception of global semantic context. This module comprises two parallel components: a wide-domain context aggregation module and a local detail enhancement module, which generate enhanced feature representations with wide-domain perception and local fidelity through channel-dimensional fusion. This design effectively bridges the gap between local perception and global understanding in traditional biomimetic architectures, significantly improving the understanding of complex highway crack scenes and ensuring the model's efficiency and robustness in UAV image processing. The overall structure of this method is as follows: Figure 3 As shown.

[0079] Preferably, the global receptive field module includes wide-domain context aggregation and local detail enhancement. After processing the feature map, it is concatenated in the channel dimension and fused and output through 1×1 convolution to generate the enhanced feature.

[0080] The wide-domain context aggregation includes at least four dilated convolution parallel branches, each of which employs dilated convolution with different dilation rates to capture multi-scale contextual information, with the set of dilation rates being {1, 2, 4, 8}.

[0081] S23. Implementation of Wide-Domain Context Aggregation Based on Global Receptive Field Modeling

[0082] The wide-domain context aggregation module is designed to address the long-distance extension characteristics of cracks, employing a multi-branch dilated convolutional architecture to capture multi-scale spatial context information. Given an input feature map... This module processes data through four parallel branches, each using a different dilation rate. The dilated convolution expands the receptive field exponentially:

[0083]

[0084] in, Indicates the expansion rate 3×3 dilated convolution operation, It is a nonlinear activation function. Different expansion rates allow each branch to perceive spatial dependencies at different scales: the small expansion rate branch ( It captures local continuity, making it suitable for detecting short, dense crack segments; large expansion rate branches ( Therefore, perceiving long-distance structural relationships is crucial for modeling the overall structure of long-distance extended cracks.

[0085] To enhance the adaptive fusion capability of features at various scales, a channel attention mechanism is introduced to calculate the weights of each branch:

[0086]

[0087] in, Indicates the first The adaptive weights of the dilated convolutional branch, with values ​​ranging from [0,1], are used to adjust the weights of the dilated convolutional branch. The output feature maps of each branch are weighted; Indicates the first Parallel branches (expansion rate) Feature map after dilated convolution; This means that the average value of the feature map in the spatial dimension is calculated to obtain a channel statistical vector; This represents the Sigmoid activation function, used to compress channel statistics to the [0,1] interval; the numerator is the... The exponentially normalized score of each branch is calculated, with the denominator being the total score of the four parallel branches. This ensures that the sum of the weights of all branches is 1, achieving normalized weighted fusion.

[0088] in, This indicates a global average pooling operation. The function maps channel statistics to the [0,1] interval, forming normalized branch weights. The features from each branch are then weighted and fused to obtain a wide-domain context representation.

[0089]

[0090] in, This represents the fused feature map output by the wide-domain context aggregation module. For the first The adaptive weights of the dilated convolutional branch, with values ​​ranging from [0,1], are used to adjust the weights of the dilated convolutional branch. The output feature maps of each branch are weighted; Indicates the first Parallel branches (expansion rate) Feature map after dilated convolution; This indicates that the feature maps of the four parallel branches are weighted and summed according to the channel dimension to achieve adaptive fusion of multi-scale features.

[0091] This module effectively captures the structural characteristics of cracks at different spatial scales through a parallel architecture of multi-scale dilated convolutions, making it particularly suitable for handling a full spectrum of targets, from tiny surface texture cracks to broad structural cracks. This design not only overcomes the limited receptive field of traditional convolutions but also significantly improves the ability to model complex crack topologies by dynamically adjusting the contribution of features at each scale for different scenarios through an adaptive weight allocation mechanism.

[0092] S24. Implementation of Local Detail Enhancement

[0093] The local detail enhancement module focuses on preserving crack edge details and high-frequency features, preventing feature smoothing and information loss issues that occur during global modeling. Considering that highway cracks often exhibit elongated structures with a width of a few pixels, this module employs a lightweight convolution and residual learning strategy to ensure the integrity of high-frequency details.

[0094] First, we reduce the dimensionality of the input features using 1×1 convolutions to decrease computational complexity.

[0095]

[0096] in, This represents the feature map after dimensionality reduction via 1×1 convolution. This represents the original feature map input to the local detail enhancement module. This represents a convolution operation with a kernel size of 1×1, used to compress or expand the channel dimension without changing the spatial resolution, thereby reducing the computational complexity of subsequent operations.

[0097] Subsequently, two consecutive 3×3 depthwise separable convolutions are applied to extract local structural features:

[0098]

[0099] in, This indicates a depthwise separable convolution operation, which significantly reduces the number of parameters while maintaining feature extraction capabilities. To enhance sensitivity to crack edges, an edge enhancement mechanism is introduced:

[0100]

[0101]

[0102] in, , , For learnable scaling factors, and These are edge detection operators, This is the generated edge enhancement map. This mechanism explicitly enhances crack boundary features, improving the network's ability to perceive slender tubular structures.

[0103] Finally, the enhanced features are fused with the original input through residual connections to ensure smooth information flow.

[0104]

[0105] in, This represents the feature map output by the local detail enhancement module. This represents the feature map after edge enhancement mechanism processing (i.e., the output after processing by the two consecutive depthwise separable convolutions and edge enhancement operators mentioned above). This represents the original feature map input to this module. It is a learnable scalar parameter used to dynamically adjust the original features. The proportion of contribution to the final output should be considered to avoid over-enhancement that could lead to loss of detail or noise amplification.

[0106] The outputs of the wide-domain context aggregation module and the local detail enhancement module are fused with a 1×1 convolution through channel concatenation to generate the final global-local collaborative representation:

[0107]

[0108] in, This represents the global-local collaborative feature map that is ultimately output by the global receptive field module. This represents the output feature map of the wide-domain context aggregation module. This represents the output feature map of the local detail enhancement module. This indicates that the two are concatenated along the channel dimension to obtain a feature map with the same number of channels as the sum of the two. This represents a convolution operation with a kernel size of 1×1, used to reduce the dimensionality of the concatenated high-dimensional features and perform nonlinear fusion to generate the final global-local collaborative representation.

[0109] This fusion strategy retains the high-frequency edge details of cracks extracted by the local detail enhancement module, and integrates the long-range structural context captured by the wide-domain context aggregation module, thus achieving a comprehensive and accurate feature representation of highway cracks.

[0110] S3. The edge features and the enhancement features are upsampled step by step by the decoder to reconstruct the image spatial resolution, and at at least one decoding level, the decoding features of the previous level are fused with the enhancement features output by the global receptive field module of the corresponding level; based on the fused decoding features of all levels, a multi-scale highway crack segmentation map is generated.

[0111] This embodiment is not a simple combination of existing modules, but rather achieves two core innovations from the perspective of synergistic integration of biological mechanism inspiration and engineering optimization:

[0112] 1. The “bionic edge perception module” designed in this embodiment simulates the SCD mechanism through dual-path contrast enhancement (Sobel+Canny) and couples spatial attention weight generation. It dynamically fuses complementary edge responses at the feature level, significantly improving the sensitivity to weak, broken or blurred crack edges, while adaptively suppressing background interference.

[0113] 2. The “global receptive field module” constructed in this embodiment breaks through the local limitations of traditional biomimetic models. It innovatively combines multi-branch dilated convolution (wide-domain context aggregation) with lightweight local detail enhancement pathways to simultaneously model long-range semantic dependencies and high-frequency edge details with low computational overhead, thereby achieving joint representation of the fine local structure and global spatial distribution of cracks.

[0114] This embodiment fundamentally solves the contradiction between biomimetic perception's "seeing clearly but not far" and global modeling's "seeing far but not clearly." Without significantly increasing model complexity, it significantly improves the accuracy, robustness, and real-time performance of UAV highway crack detection, providing a deployable and highly reliable new paradigm for intelligent infrastructure inspection in complex scenarios.

[0115] To verify the beneficial effects of this embodiment, the detection method provided in this embodiment was run after setting the evaluation parameters, and the experimental results are as follows.

[0116] The dataset used in this study is the RCI (Aerial RoadCrack Dataset), specifically designed for highway crack detection tasks. This dataset contains 1880 aerial images, collected in the field by a DJI Inspire 1 drone, covering diverse terrains and complex environmental conditions, including 20 images simulating snowy scenes to enhance the model's adaptability to extreme weather. The dataset is divided into three parts according to standard: 1316 training images, 376 validation images, and 188 test images, ensuring the evaluation results have statistical significance and generalization ability. Its core advantage lies in its comprehensive coverage of varied terrain features (such as slopes, farmland, and urban-rural fringe areas) and different lighting and weather conditions, making it not only suitable for crack identification and classification tasks but also an ideal benchmark platform for verifying the algorithm's performance in terrain perception and environmental adaptability. This dataset is particularly suitable for developing and validating robust crack detection models for real-world field inspection scenarios.

[0117] The training parameters for this experiment are shown in Table 1. The training was conducted on a workstation equipped with an Intel(R) Xeon(R) Silver 4210R CPU@2.40GHz, 64GB of RAM, and an NVIDIA GeForce RTX 3080 GPU. The software environment was based on Windows 10, using Python 3.8.18 and CUDA 12.2 to accelerate deep learning computations. Model training used the stochastic gradient descent (SGD) optimizer for 200 iterations, with a batch size of 8 and an initial learning rate fixed at 0.01.

[0118] Table 1 Training Parameter Settings

[0119]

[0120] To ensure a fair and comprehensive comparison between different methods and a thorough evaluation of model performance, multiple metrics are used. Precision is defined as the proportion of true positives (TP) out of all predicted positives (TP+FP), used to measure the accuracy of the prediction results; Recall is defined as the proportion of true positives (TP) out of all actual positives (TP+FN), used to measure the model's coverage of positive samples; Mean AP (mAP) is the average precision across M classes. The average value is an important indicator for comprehensively evaluating the performance of models in multi-class object detection tasks. These three formulas together form the basic framework for analyzing the performance of object detection models.

[0121]

[0122]

[0123]

[0124] As shown in Table 2, the method in this embodiment performs optimally in UAV-based highway crack detection tasks: mAP@0.5 reaches 89.3%, Recall is 83.7%, and Precision is 87.5%, all outperforming mainstream models such as YOLOv5, YOLOv7, and YOLOv11. This indicates that the proposed method has stronger detection and recognition capabilities for small and irregular cracks in complex road surface environments, balancing high accuracy and high recall, making it more suitable for automated crack detection needs in practical road inspection scenarios.

[0125] Table 2 Comparison of Experimental Results

[0126] mAP@0.5% Recall% Precision% YOLOv5 82.1 78.4 84.6 YOLOv7 84.5 80.2 86.3 YOLOv11 86.8 82.0 87.1 This embodiment 89.3 83.7 87.5

[0127] This method can accurately detect and identify road cracks captured by drones in a variety of complex scenarios: Figure 4 It demonstrates high-confidence identification of small, low-contrast cracks (confidence level 0.74). Figure 5 Successfully captured multiple longitudinal and transverse cracks against the backdrop of urban roads, maintaining stable detection even under shadows and road marking interference (confidence level 0.83–0.85). Figure 6 Even under harsh conditions such as snow cover and uneven lighting, the method still accurately outlines penetrating cracks (confidence level 0.81–0.84). The visualization results fully verify the robustness and generalization ability of this method in real road inspection scenarios.

[0128] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting road surface cracks based on edge perception and global receptive field modeling, characterized in that, Includes the following steps: S1. Input the image to be detected into the encoder-decoder network, and extract multi-scale features from the input image through the encoder to obtain feature maps at different levels; S2. Input the feature map of at least one layer in the encoder into the bionic edge perception module to enhance the edge features; The feature maps of at least one other layer are respectively input into the corresponding global receptive field module, and the context aggregation and detail enhancement are fused together to output the enhanced features. The biomimetic edge perception module includes dual-path contrast enhancement and spatial attention weight generation; The dual-path contrast enhancement is used to enhance the feature map, and the process includes: The feature map is extracted in parallel using two-path features, wherein the first path extracts the edge features and the second path extracts global statistical features; The edge features and the global statistical features are concatenated along the channel dimension to obtain the fused features; After channel expansion, the fused features are added element-wise to the original input feature map in the spatial domain to output the enhanced edge features. The global receptive field module includes wide-domain context aggregation and local detail enhancement. After processing the feature map, it concatenates it in the channel dimension and then fuses and outputs it through a 1×1 convolution to generate the enhanced feature. The wide-domain context aggregation includes dilated convolution parallel branches to capture multi-scale contextual information; S3. The edge features and the enhancement features are upsampled step by step by the decoder to reconstruct the image spatial resolution, and at at least one decoding level, the decoding features of the previous level are fused with the enhancement features output by the global receptive field module of the corresponding level; based on the fused decoding features of all levels, a highway crack segmentation map is generated.

2. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The first path extracts complementary edge features by applying the Sobel and Canny operators in parallel to the feature map.

3. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The second path extracts the global statistical features from the feature map through global average pooling and global max pooling.

4. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The spatial attention weight generation takes the output of the dual-path contrast enhancement as input, generates an initial attention map through a lightweight convolutional layer and a non-linear activation function, and then normalizes it to obtain a spatial weight map.

5. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The local detail enhancement described in step S2 is achieved through lightweight convolution operations to preserve the local detail information of the feature map.

6. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The wide-domain context aggregation includes at least four parallel branches, each of which employs dilated convolutions with different dilation rates, the set of dilation rates being {1, 2, 4, 8}.

7. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The wide-domain context aggregation also uses a channel attention mechanism to adaptively generate fusion weights for the output features of each parallel branch.

8. The road surface crack detection method based on edge perception and global receptive field modeling according to claim 1, characterized in that, The process of enhancing local details includes: Dimensionality reduction is performed on the feature map; Perform continuous depthwise separable convolutions on the reduced feature map; The result of the depthwise separable convolution is added to the feature map of the original input via residual connections.

Images