Feature extraction method for road disease detection and related device

Abstract

This invention provides a feature extraction method and related equipment for road defect detection. The feature extraction method includes: acquiring the original feature map of the road image to be detected output by a target detection backbone network, and obtaining a multi-scale feature map based on the original feature map; performing channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map; wherein, during the channel-level feature recalibration process, adaptively amplifying channel weights sensitive to road defects and suppressing channel weights sensitive to background noise; filtering the background in the channel feature map to obtain defect candidate regions; performing word-level self-attention calculation on the defect candidate regions to obtain a purified feature map; and fusing the original feature map and the purified feature map to obtain a target feature map for road defect detection. The technical solution provided by this invention can generate high-quality target feature maps that are information-rich, context-aware, and well-suppressed by noise.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and road maintenance technology, and in particular to a feature extraction method and related equipment for road defect detection. Background Technology

[0002] In complex road surface inspection scenarios, road defects such as cracks and small potholes have strong spatial sparsity. The features extracted by the backbone network often contain a large amount of useless asphalt background and light interference information, resulting in low feature purity participating in downstream fusion, which seriously affects the detection accuracy and leads to insufficient defect detection accuracy. Summary of the Invention

[0003] This invention provides a feature extraction method and related equipment for road defect detection, in order to overcome the deficiencies in the prior art.

[0004] This invention provides a feature extraction method for road defect detection, comprising: Obtain the original feature map of the road image to be detected output by the target detection backbone network, and obtain a multi-scale feature map based on the original feature map; The multi-scale feature map is recalibrated at the channel level to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights that are sensitive to road defects are adaptively amplified, while the channel weights that are sensitive to background noise are suppressed. The background in the channel feature map is filtered to obtain candidate disease regions; The candidate disease regions are subjected to word-level self-attention calculation to obtain purified feature maps; The original feature map is fused with the purified feature map to obtain a target feature map for road defect detection.

[0005] According to the feature extraction method for road defect detection provided by the present invention, the step of performing channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map includes: The multi-scale feature map is subjected to global average pooling along the spatial dimension to obtain channel-level statistical vectors; The channel-level statistical vectors are input into a multilayer perceptron and combined with the ReLU activation function and the Sigmoid activation function to calculate the channel weights. The channel weights are multiplied element-wise with the multi-scale feature map to obtain the channel feature map.

[0006] According to the feature extraction method for road defect detection provided by the present invention, the step of filtering the background in the channel feature map to obtain defect candidate regions includes: The channel feature map is divided into multiple non-overlapping macro regions in the spatial dimension; Calculate the first query vector and the first key vector for each of the macro regions; Based on the first query vector and the first key vector, calculate the affinity matrix for each of the macroscopic regions; Apply the Top-K routing algorithm to each row of each affinity matrix to prune the connections of the background region and generate a routing index matrix; The background in the channel feature map is filtered based on the routing index matrix to obtain the disease candidate region.

[0007] According to the feature extraction method for road defect detection provided by the present invention, the step of performing word-level self-attention calculation on the defect candidate region to obtain a purified feature map includes: Calculate the second query vector, second key vector, and value vector of the terms in the disease candidate region; Based on the second query vector, the second key vector, and the value vector, word-level self-attention calculation is performed on the disease candidate region to obtain the purified feature map.

[0008] According to the feature extraction method for road defect detection provided by the present invention, the step of obtaining a multi-scale feature map based on the original feature map includes: The multi-scale feature map is composed of feature maps output by multiple target feature layers of the target detection backbone network in the original feature map.

[0009] According to a feature extraction method for road defect detection provided by the present invention, after fusing the original feature map with the purified feature map to obtain a target feature map for road defect detection, the method further includes: Based on the target feature map, road defects are detected to obtain road defect detection results; wherein, the road defect detection results include the type, location and confidence level of the road defects.

[0010] The present invention also provides a feature extraction device for road defect detection, comprising: The acquisition module is configured to acquire the original feature map of the road image to be detected output by the target detection backbone network, and obtain a multi-scale feature map based on the original feature map; The feature recalibration module is configured to perform channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights sensitive to road defects are adaptively amplified, while the channel weights sensitive to background noise are suppressed. The filtering module is configured to filter the background in the channel feature map to obtain candidate disease regions; The calculation module is configured to perform word-level self-attention calculation on the disease candidate region to obtain a purified feature map; The fusion module is configured to fuse the original feature map with the purified feature map to obtain a target feature map for road defect detection.

[0011] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the feature extraction method for road defect detection as described above.

[0012] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the feature extraction method for road defect detection as described above.

[0013] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the feature extraction method for road defect detection as described above.

[0014] The feature extraction method and related equipment for road defect detection provided by this invention achieve a feature enhancement process from coarse to fine and from local to global through multi-scale feature generation, channel-level feature recalibration, background filtering, word-level self-attention calculation, and final feature fusion. First, the multi-scale design ensures robustness to defects of different sizes. Second, channel-level feature recalibration effectively enhances road defect-related features and suppresses background noise. Next, the combination of background filtering and self-attention mechanisms effectively reduces computational load while accurately capturing long-range dependencies within road defects, significantly improving feature discriminative power. Finally, fusion of original features compensates for any loss of detail information during processing. The technical solution provided by this invention can generate high-quality target feature maps that are information-rich, context-aware, and well-suppressed by noise, thereby significantly improving the accuracy and robustness of subsequent road defect detection tasks. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the feature extraction method for road defect detection provided by the present invention.

[0017] Figure 2 This is a simplified schematic diagram of the feature extraction method for road defect detection provided by the present invention.

[0018] Figure 3 This is a schematic diagram of the road defect detection results provided by the present invention.

[0019] Figure 4 This is a schematic diagram of the feature extraction device for road defect detection provided by the present invention.

[0020] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] The data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are all information and data that have been fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0023] Figure 1 This is a flowchart illustrating a feature extraction method for road defect detection according to an exemplary embodiment. Figure 1 As shown in an exemplary embodiment, the feature extraction method for road defect detection includes steps 110 to 150, which are described in detail below.

[0024] Step 110: Obtain the original feature map of the road image to be detected output by the target detection backbone network, and obtain a multi-scale feature map based on the original feature map.

[0025] In this embodiment of the invention, images of the road to be inspected are collected using equipment such as vehicle-mounted cameras, drone aerial photography equipment, or specialized road inspection vehicles.

[0026] A pre-built road defect detection network is constructed, comprising a target detection backbone network with a Spatial Tuning Adapter (STA) module, a Bi-Fusion fusion node, and a detection head. A feature pre-refinement module (Coordinate Attention-Bi-directional-path Refinement Attention, CA-BRA) is set before the Bi-Fusion fusion node of the STA module. Figure 2 As shown, the feature pre-purification module includes a series of dynamic channel attention units and a dual-layer routing attention unit. It adopts a series of dual filtering mechanisms of "macro-filtering first, then micro-focusing". First, the channel dimension is purified through dynamic channel attention, and then the fine-grained feature focusing of the background area and the disease area is achieved through dual-layer routing attention.

[0027] The object detection backbone network is a fundamental network structure used for feature extraction in deep learning. This backbone network can be any mature convolutional neural network or Transformer architecture. The acquired road image to be detected is input into the object detection backbone network, which performs deep feature extraction, capturing information such as texture, contours, and color to obtain a raw feature map. This raw feature map contains a preliminary encoding of the image content, including information about road defects such as cracks and potholes, as well as a large amount of background information, such as road markings, tire tracks, water stains, and shadows.

[0028] Because road defects (such as cracks, potholes, and network cracks) vary in size and shape, a single-scale feature map may be insufficient to effectively represent both large and small defects simultaneously. Therefore, multi-scale feature maps are derived from the original feature maps. Specifically, multi-scale feature maps are obtained by fusing high-level strong semantic information with low-level strong spatial information (high resolution), resulting in different resolutions.

[0029] Step 120: Perform channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map; wherein, during the channel-level feature recalibration process, adaptively amplify the channel weights that are sensitive to road defects and suppress the channel weights that are sensitive to background noise.

[0030] In this embodiment of the invention, different feature channels in the multi-scale feature map have different emphases in representing information. For example, some channels may respond more strongly to the edge texture of cracks, while others may be more sensitive to the shadows or depth information of potholes, and a large number of channels may mainly respond to background noise. In order to enhance the road defect-related features and suppress irrelevant noise, channel-level feature recalibration is performed on the multi-scale feature map to obtain a channel feature map.

[0031] Specifically, channel-level feature recalibration involves learning a channel weight for each channel of the multi-scale feature map, where the channel weight represents the importance of that channel. During the recalibration process, channel weights sensitive to road defects are adaptively amplified, while channel weights sensitive to background noise are suppressed, dynamically adjusting the contribution of each channel.

[0032] Step 130: Filter the background in the channel feature map to obtain candidate disease regions.

[0033] In this embodiment of the invention, although the road defect features are enhanced through channel-level feature recalibration, a large number of background areas still exist in the channel feature map. To reduce the computational overhead of subsequent complex calculations and to further focus on areas where defects may exist, the background in the channel feature map is filtered to obtain defect candidate regions.

[0034] Step 140: Perform word-level self-attention calculation on the disease candidate region to obtain a purified feature map.

[0035] In this embodiment of the invention, road defects, especially cracks, typically exhibit long-range dependencies, meaning that a part of the defect is structurally related to another part at a distance. Traditional convolutional operations, due to their limited receptive field, struggle to capture such long-range dependencies. This embodiment of the invention addresses this issue by introducing a self-attention mechanism, performing token-level self-attention computation on defect candidate regions to obtain a refined feature map. A token refers to the feature vector at each spatial location within a defect candidate region.

[0036] Step 150: The original feature map and the purified feature map are fused to obtain a target feature map for road defect detection.

[0037] In this embodiment of the invention, the purified feature map possesses strong contextual relationships and semantic information, but after multiple processing steps, it may lose some fine, low-level spatial positioning information. The original feature map, however, is rich in this detailed information. Therefore, the two are fused to achieve complementary advantages.

[0038] Specifically, the methods of integration may include, but are not limited to: Element-wise addition: The two aligned feature maps are added element-wise at their corresponding positions; Channel-level concatenation: Two feature maps are concatenated along the channel dimension, and then one or more convolutional layers (such as 1x1 convolution) are added to integrate the information and adjust the number of channels.

[0039] Specifically, the purified feature map is input into the Bi-Fusion fusion node of the road disease detection basic network and fused with the original feature map to complete the feature pre-purification process.

[0040] The feature map obtained after fusion is the final target feature map. This target feature map contains both high-level semantic information about road defects with global context awareness, purified through a self-attention mechanism, and original spatial details from the backbone network that help with accurate localization. This target feature map can be directly fed into subsequent detection heads for tasks such as road defect classification.

[0041] This invention, through multi-scale feature generation, channel-level feature recalibration, background filtering, term-level self-attention computation, and final feature fusion, achieves a feature enhancement process from coarse to fine and from local to global. First, the multi-scale design ensures robustness to defects of different sizes. Second, channel-level feature recalibration effectively enhances defect-related features and suppresses background noise. Next, the combination of background filtering and self-attention mechanisms effectively reduces computational load while accurately capturing long-range dependencies within road defects, significantly improving feature discriminative power. Finally, fusing original features compensates for any loss of detail information during processing. The technical solution provided by this invention can generate high-quality target feature maps that are information-rich, context-aware, and well-suppressed by noise, thereby significantly improving the accuracy and robustness of subsequent road defect detection tasks.

[0042] In an exemplary embodiment of the present invention, the step of performing channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map includes: The multi-scale feature map is subjected to global average pooling along the spatial dimension to obtain channel-level statistical vectors; The channel-level statistical vectors are input into a multilayer perceptron and combined with the ReLU activation function and the Sigmoid activation function to calculate the channel weights. The channel weights are multiplied element-wise with the multi-scale feature map to obtain the channel feature map.

[0043] In this embodiment of the invention, the l-th layer feature map of the multi-scale feature map... Global average pooling is performed along the spatial dimension to compress it spatially, resulting in a channel-level statistical vector that includes the global receptive field. , where H, W, and C are the height, width, and number of channels of the feature map, respectively.

[0044] The channel-level statistical vectors are input into a multilayer perceptron consisting of two fully connected layers. By combining the ReLU and Sigmoid activation functions, the channel nonlinear weight coefficients for each channel are calculated, thus obtaining the channel weights. In a multilayer perceptron, the first fully connected layer is used for dimensionality reduction to decrease computation, and the second fully connected layer restores the dimensionality.

[0045] The channel weights are multiplied element-wise with the original input feature map at the channel level. The feature responses of channels with higher weights (i.e., those considered more important for detecting road defects) are amplified, while the feature responses of channels with lower weights (such as those corresponding to background noise) are suppressed, resulting in a channel feature map that has been purified. .

[0046] In an exemplary embodiment of the present invention, filtering the background in the channel feature map to obtain candidate disease regions includes: The channel feature map is divided into multiple non-overlapping macro regions in the spatial dimension; Calculate the first query vector and the first key vector for each of the macro regions; Based on the first query vector and the first key vector, calculate the affinity matrix for each of the macroscopic regions; Apply the Top-K routing algorithm to each row of each affinity matrix to prune the connections of the background region and generate a routing index matrix; The background in the channel feature map is filtered based on the routing index matrix to obtain the disease candidate region.

[0047] In this embodiment of the invention, the channel feature map after channel purification Divided in spatial dimension For each of the three non-overlapping macro-regions, calculate the first query vector at the region level. With the first key vector Through the first query vector With the first key vector Constructing an affinity matrix T denotes transpose. This is achieved through the affinity matrix. Assess the semantic relevance between regions.

[0048] Apply to each row of the affinity matrix Routing algorithm, pruning background region connections, generating routing index matrix This completes the hard filtering of the macro background.

[0049] In an exemplary embodiment of the present invention, the step of performing word-level self-attention calculation on the disease candidate region to obtain a purified feature map includes: Calculate the second query vector, second key vector, and value vector of the terms in the disease candidate region; Based on the second query vector, the second key vector, and the value vector, word-level self-attention calculation is performed on the disease candidate region to obtain the purified feature map.

[0050] In this embodiment of the invention, word-level self-attention calculation is performed only on the disease candidate regions selected by the routing index matrix to obtain the final purified feature map. The self-attention calculation expression is: ; in, This is the second query vector for word i within the disease candidate region. Let i be the second key vector and value vector. The output is the purification feature.

[0051] In an exemplary embodiment of the present invention, obtaining a multi-scale feature map based on the original feature map includes: The multi-scale feature map is composed of feature maps output by multiple target feature layers of the target detection backbone network in the original feature map.

[0052] In this embodiment of the invention, the target detection backbone network adopts the DINOv3 (Distillation with NO labels version 3) network. The DINOv3 network is a self-supervised visual model. Its core goal is to learn general, high-performance visual features without any manual annotation. These features can be directly used for various downstream tasks (such as image classification, segmentation, depth estimation, etc.) and the results are usually better than supervised pre-trained models.

[0053] Multiple feature layers are selected as target feature layers in the target detection backbone network, such as layers 5, 8, and 11 of the DINOv3 network. The feature maps output by the target feature layers are combined to form a multi-scale feature map.

[0054] Specifically, by utilizing the outputs of multiple layers at different depths (i.e. different scales) in the object detection backbone network, strong semantic information at high levels is fused with strong spatial information (high resolution) at low levels through top-down paths and lateral connections, thereby generating a series of feature maps with different resolutions, which together constitute a multi-scale feature map.

[0055] In an exemplary embodiment of the present invention, after fusing the original feature map with the purified feature map to obtain a target feature map for road defect detection, the method further includes: Based on the target feature map, road defects are detected to obtain road defect detection results; wherein, the road defect detection results include the type, location and confidence level of the road defects.

[0056] In this embodiment of the invention, the fused target feature map is fed into a detection head for road defect detection, resulting in road defect detection results that include the type, location, and confidence level of the road defects. Figure 3 As shown, the type and location of road defects are marked on the road image to be detected.

[0057] The feature extraction device for road defect detection provided by the present invention will be described below. The feature extraction device for road defect detection described below can be referred to in correspondence with the feature extraction method for road defect detection described above. It should be noted that the device provided in the following embodiments belongs to the same concept as the method provided in the above embodiments, and the specific way in which each module and unit performs its operation has been described in detail in the method embodiments, and will not be repeated here.

[0058] In one exemplary embodiment of the present invention, please refer to Figure 4 , Figure 4 This is a feature extraction device for road defect detection according to an exemplary embodiment, comprising the following modules.

[0059] The acquisition module 410 is configured to acquire the original feature map of the road image to be detected output by the target detection backbone network, and obtain a multi-scale feature map based on the original feature map. The feature recalibration module 420 is configured to perform channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights sensitive to road defects are adaptively amplified, while the channel weights sensitive to background noise are suppressed. Filtering module 430 is configured to filter the background in the channel feature map to obtain disease candidate regions; The calculation module 440 is configured to perform word-level self-attention calculation on the disease candidate region to obtain a purified feature map; The fusion module 450 is configured to fuse the original feature map with the purified feature map to obtain a target feature map for road defect detection.

[0060] In an exemplary embodiment of the present invention, the feature recalibration module 420 includes: The pooling submodule is configured to perform global average pooling on the multi-scale feature map along the spatial dimension to obtain a channel-level statistical vector. The first calculation submodule is configured to input the channel-level statistical vector into the multilayer perceptron and calculate the channel weights by combining the ReLU activation function and the Sigmoid activation function. The element-wise multiplication submodule is configured to multiply the channel weights element-wise with the multi-scale feature map to obtain the channel feature map.

[0061] In an exemplary embodiment of the present invention, the filtering module 430 includes: The sub-module is configured to divide the channel feature map into multiple non-overlapping macro regions in the spatial dimension; The second calculation submodule is configured to calculate the first query vector and the first key vector for each of the macro regions; The third calculation submodule is configured to calculate the affinity matrix of each of the macroscopic regions based on the first query vector and the first key vector. The application submodule is configured to apply the Top-K routing algorithm to each row of the affinity matrix, prune the connections of the background region, and generate a routing index matrix. The filtering submodule is configured to filter the background in the channel feature map based on the routing index matrix to obtain the disease candidate region.

[0062] In an exemplary embodiment of the present invention, the computing module 440 includes: The fourth calculation submodule is configured to calculate the second query vector, the second key vector, and the value vector of the terms in the disease candidate region; The fifth calculation submodule is configured to perform word-level self-attention calculation on the disease candidate region based on the second query vector, the second key vector, and the value vector to obtain the purified feature map.

[0063] In an exemplary embodiment of the present invention, the acquisition module 410 includes: The submodule is configured to combine the feature maps output by multiple target feature layers of the target detection backbone network in the original feature map to form the multi-scale feature map.

[0064] In an exemplary embodiment of the present invention, the feature extraction device for road defect detection further includes: The road defect detection module is configured to perform road defect detection based on the target feature map to obtain road defect detection results; wherein, the road defect detection results include the type, location and confidence level of the road defects.

[0065] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a feature extraction method for road defect detection. This method includes: acquiring the original feature map of the road image to be detected output by the target detection backbone network, and obtaining a multi-scale feature map based on the original feature map. The multi-scale feature map is recalibrated at the channel level to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights that are sensitive to road defects are adaptively amplified, while the channel weights that are sensitive to background noise are suppressed. The background in the channel feature map is filtered to obtain candidate disease regions; The candidate disease regions are subjected to word-level self-attention calculation to obtain purified feature maps; The original feature map is fused with the purified feature map to obtain a target feature map for road defect detection.

[0066] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute the feature extraction method for road defect detection provided by the above methods, the method including: acquiring the original feature map of the road image to be detected output by the target detection backbone network, and obtaining a multi-scale feature map based on the original feature map; The multi-scale feature map is recalibrated at the channel level to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights that are sensitive to road defects are adaptively amplified, while the channel weights that are sensitive to background noise are suppressed. The background in the channel feature map is filtered to obtain candidate disease regions; The candidate disease regions are subjected to word-level self-attention calculation to obtain purified feature maps; The original feature map is fused with the purified feature map to obtain a target feature map for road defect detection.

[0068] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a feature extraction method for road defect detection provided by the above methods, the method comprising: acquiring an original feature map of a road image to be detected output by a target detection backbone network, and obtaining a multi-scale feature map based on the original feature map; The multi-scale feature map is recalibrated at the channel level to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights that are sensitive to road defects are adaptively amplified, while the channel weights that are sensitive to background noise are suppressed. The background in the channel feature map is filtered to obtain candidate disease regions; The candidate disease regions are subjected to word-level self-attention calculation to obtain purified feature maps; The original feature map is fused with the purified feature map to obtain a target feature map for road defect detection.

[0069] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0070] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A feature extraction method for road defect detection, characterized in that, include: Obtain the original feature map of the road image to be detected output by the target detection backbone network, and obtain a multi-scale feature map based on the original feature map; The multi-scale feature map is recalibrated at the channel level to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights that are sensitive to road defects are adaptively amplified, while the channel weights that are sensitive to background noise are suppressed. The background in the channel feature map is filtered to obtain candidate disease regions; The candidate disease regions are subjected to word-level self-attention calculation to obtain purified feature maps; The original feature map is fused with the purified feature map to obtain a target feature map for road defect detection.

2. The feature extraction method for road defect detection according to claim 1, characterized in that, The process of performing channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map includes: The multi-scale feature map is subjected to global average pooling along the spatial dimension to obtain channel-level statistical vectors; The channel-level statistical vectors are input into a multilayer perceptron and combined with the ReLU activation function and the Sigmoid activation function to calculate the channel weights. The channel weights are multiplied element-wise with the multi-scale feature map to obtain the channel feature map.

3. The feature extraction method for road defect detection according to claim 1, characterized in that, The step of filtering the background in the channel feature map to obtain candidate disease regions includes: The channel feature map is divided into multiple non-overlapping macro regions in the spatial dimension; Calculate the first query vector and the first key vector for each of the macro regions; Based on the first query vector and the first key vector, calculate the affinity matrix for each of the macroscopic regions; Apply the Top-K routing algorithm to each row of each affinity matrix to prune the connections of the background region and generate a routing index matrix; The background in the channel feature map is filtered based on the routing index matrix to obtain the disease candidate region.

4. The feature extraction method for road defect detection according to claim 1, characterized in that, The step of performing word-level self-attention calculation on the candidate disease regions to obtain a purified feature map includes: Calculate the second query vector, second key vector, and value vector of the terms in the disease candidate region; Based on the second query vector, the second key vector, and the value vector, word-level self-attention calculation is performed on the disease candidate region to obtain the purified feature map.

5. The feature extraction method for road defect detection according to any one of claims 1 to 4, characterized in that, The process of obtaining a multi-scale feature map based on the original feature map includes: The multi-scale feature map is composed of feature maps output by multiple target feature layers of the target detection backbone network in the original feature map.

6. The feature extraction method for road defect detection according to any one of claims 1 to 4, characterized in that, After fusing the original feature map with the purified feature map to obtain the target feature map for road defect detection, the method further includes: Based on the target feature map, road defects are detected to obtain road defect detection results; wherein, the road defect detection results include the type, location and confidence level of the road defects.

7. A feature extraction device for road defect detection, characterized in that, include: The acquisition module is configured to acquire the original feature map of the road image to be detected output by the target detection backbone network, and obtain a multi-scale feature map based on the original feature map; The feature recalibration module is configured to perform channel-level feature recalibration on the multi-scale feature map to obtain a channel feature map; wherein, during the channel-level feature recalibration process, the channel weights sensitive to road defects are adaptively amplified, while the channel weights sensitive to background noise are suppressed. The filtering module is configured to filter the background in the channel feature map to obtain candidate disease regions; The calculation module is configured to perform word-level self-attention calculation on the disease candidate region to obtain a purified feature map; The fusion module is configured to fuse the original feature map with the purified feature map to obtain a target feature map for road defect detection.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the feature extraction method for road defect detection as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the feature extraction method for road defect detection as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the feature extraction method for road defect detection as described in any one of claims 1 to 6.

Images