Background suppression method and system based on unmanned aerial vehicle inspection

Abstract

The application discloses a background suppression method and system based on unmanned aerial vehicle inspection, and belongs to the technical field of intelligent detection, and comprises the following steps: S1, standard crack feature construction; S2, road local feature extraction; S3, cosine similarity calculation; S4, background suppression enhancement; S5, multi-scale feature extraction; and S6, crack detection output. Through the above manner, the standard crack feature vector is constructed, the unmanned aerial vehicle inspection image is divided into blocks, and then cosine similarity matching is performed on the standard crack features, so that the non-disease background such as shadows, water stains and road textures in the image can be suppressed in a targeted manner before detection, and the regions with relatively high similarity to the crack features are reserved and enhanced, compared with the mode of only relying on the detection network to learn the background difference itself.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, specifically to a background suppression method and system based on unmanned aerial vehicle (UAV) inspection. Background Technology

[0002] Road inspection is a crucial part of road maintenance and traffic safety management. Inspection targets typically include road surface defects such as cracks, potholes, and other damage. With the development of drone platforms, high-definition imaging equipment, and deep learning vision algorithms, road inspection is gradually evolving from manual on-site checks towards automation, intelligence, and high-frequency operation. Drones can rapidly acquire road images over a large area, while computer vision technologies such as target detection, semantic segmentation, and feature enhancement can be used to identify defects such as cracks from inspection images.

[0003] Existing YOLOv11 object detection networks typically employ an end-to-end single-stage detection process to complete road defect identification tasks. Specifically, the input road image is fed into the network backbone feature extraction module, which extracts low-level texture features and high-level semantic features of the image step by step through a multi-layer convolutional structure. Subsequently, the feature maps are fused at multiple scales in the neck network through a feature pyramid structure to output the final road defect detection result.

[0004] However, the YOLOv11 detection network still has certain limitations in real-world complex road inspection scenarios: it lacks the ability to represent the features of targets such as small cracks, which can easily lead to missed detections; at the same time, it is prone to false detections in complex backgrounds such as shadows, water stains, and road surface textures, which reduces the reliability of the detection results. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a background suppression method and system based on UAV inspection.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Background suppression methods based on drone inspection include the following steps: S1. Mark the target bounding box on the crack image, extract the local features of the crack, and perform pooling and fusion processing on multiple local crack features to form standard crack features; S2. Acquire images of the road to be detected, divide them into multiple image blocks of uniform size, and extract local road features block by block through convolution. S3. Obtain local road features and standard crack features, calculate the similarity between features using cosine similarity, and generate a similarity matrix based on the row and column positions of the local road feature image blocks in the road image to be detected. S4. Based on the similarity matrix, perform background suppression and crack feature enhancement on each image block region in the road image to be detected, and output a background suppression and enhancement map. This includes the following steps: (1) Calculate the background suppression weights; (2) Use background suppression weights to weight and enhance the original image patch features to obtain the enhanced features after background suppression; (3) Obtain the background suppression weight map; (4) The background suppression weight map is used to weight the image to be detected and output the background suppression enhancement map; S5. Input the background suppression enhancement map into the backbone feature extraction part of the YOLOv11 detection network, and perform convolutional normalization preprocessing, deep feature extraction, C2PSA attention enhancement and spatial pyramid pooling SPPF processing in sequence to obtain multi-scale basic feature maps. S6. Input the multi-scale basic feature map into the feature fusion part of the YOLOv11 detection network. Use upsampling and stitching to fuse deep high semantic features and shallow high resolution detail features to obtain fused multi-scale features. Then, send the fused multi-scale features into the detection head to output the crack's location coordinates, size information and confidence level to complete the final identification.

[0007] Furthermore, step S1 specifically includes: Multiple clearly labeled images of road cracks from drones were selected as input images, and the crack targets in the input images were labeled with bounding boxes. Obtain the bounding box annotations of the input image, and extract local features of the cracks through convolution, normalization, and activation functions; Spatial pooling, normalization, and mean fusion are performed on the local crack features of multiple crack images to form stable standard crack features.

[0008] Furthermore, step S2 specifically includes: Images of the road to be inspected are collected by drones. First, the images are sized uniformly and then normalized. The preprocessed road image to be detected is evenly divided into multiple image blocks of a fixed size; For each image block, local road features are extracted block by block through convolution.

[0009] Furthermore, step S3 specifically includes: Cosine similarity is calculated between local road features and standard crack features to measure the closeness between each image patch and the standard crack features. Based on the original spatial location of each road local feature image block in the image of the road to be detected, the corresponding similarity value is mapped back to the block grid to generate a similarity matrix that is consistent with the spatial layout of the image blocks.

[0010] Furthermore, in step (1), the background suppression weight calculation formula is as follows: in, The background suppression weights for the i-th block are... Here, is the Sigmoid function, T is the similarity threshold, and k is a coefficient that adjusts the suppression strength.

[0011] Furthermore, in step (2), let the original image patch features be... The formula for calculating the enhanced features after background suppression is: in, This is an enhanced feature after background suppression. The original block features, For feature enhancement coefficients, is the background suppression weight for the i-th block.

[0012] Furthermore, step S5 specifically includes the following steps: S501. The input background suppression enhancement map first passes through basic units consisting of convolutional layers, batch normalization layers, and SiLU activation functions to form the initial feature mapping and standardization of the image. S502. The C3k2 module performs in-depth feature extraction and channel interaction processing on the preliminary features of the executed image to obtain the extracted features. S503. Input the extracted features into the Spatial Pyramid Pooling (SPPF) module, and aggregate contextual information at different scales through continuous pooling operations to obtain multi-scale contextual features. S504. Input the multi-scale contextual features into the C2PSA attention module to enhance the features in both channel and spatial dimensions, obtain attention-enhanced features, and finally output the multi-scale basic feature map.

[0013] Furthermore, step S6 includes the following steps: S601. Upsample the multi-scale basic feature map step by step and merge it with the corresponding shallow features; then merge it with the high-level features through step-by-step downsampling. After integration by the C3k2 module, the fused multi-scale features are obtained. S602. The fused multi-scale features are fed into the detection head. The detection head predicts the bounding box position, class confidence and target confidence of the target through parallel branches, and outputs the location coordinates, size information and confidence score of the crack.

[0014] A background suppression system based on unmanned aerial vehicle (UAV) inspection, the system comprising: Crack feature construction module: Input crack images, mark target boxes, extract local crack features, and perform pooling and fusion processing on the local crack features of multiple crack images to obtain standard crack features; Road local feature extraction module: Input the road image to be detected, divide the image into multiple image blocks with uniform size, and extract the road local features corresponding to the image blocks through convolution feature extraction processing; Cosine similarity calculation module: acquires local road features and standard crack features, calculates the similarity between each road image block and the standard crack features through cosine similarity calculation, and generates a similarity matrix; Background suppression enhancement module: Input similarity matrix and local road features, calculate background suppression weights, perform weighted enhancement on the original image patch features based on the background suppression weights, obtain the enhanced features after background suppression, restore the size of the background suppression weights to obtain the background suppression weight map, perform weighted processing on the image to be detected based on the background suppression weight map, and output the background suppression enhancement map; YOLOv11 Detection Network: The background suppression enhancement map is input into the YOLOv11 detection network. Through convolutional normalization preprocessing, C3k2 deep feature extraction, C2PSA attention enhancement, and SPPF spatial pyramid pooling, a multi-scale basic feature map with both local detail information and global semantic information is obtained. The multi-scale basic feature map is input into the feature fusion part of the YOLOv11 detection network. Multi-scale fused features are generated through upsampling and Concat feature fusion. The detection head is then used to predict the location, size, and confidence of the crack target to obtain the crack detection result.

[0015] Compared with the prior art, the beneficial effects of this invention are as follows: 1. This invention constructs a standard crack feature vector and performs cosine similarity matching between the UAV inspection image and the standard crack feature vector after segmentation. This enables targeted suppression of non-disease backgrounds such as shadows, water stains, and road textures in the image before detection. At the same time, it preserves and enhances areas with high similarity to crack features. Compared with the method that relies solely on the detection network to learn background differences, this scheme can more directly highlight the crack target features, reduce the interference of complex backgrounds on the detection results, thereby reducing false detections due to backgrounds, and is conducive to preserving key disease information such as small cracks and complex cracks. 2. This invention combines the background suppression enhancement module with the YOLOv11 detection network, multi-scale feature extraction, attention enhancement, and feature fusion, enabling the model to not only focus on the local texture and edge details of cracks, but also to use deep semantic information to complete localization and recognition. At the same time, through closed-loop optimization of threshold and enhancement coefficient, it improves the adaptability of different road scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the optimized detection model of the present invention; Figure 2 This is a flowchart of the background suppression enhancement module of the present invention; Figure 3 The output result diagram of the optimized detection model of this invention is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present 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 the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Background suppression method based on UAV inspection, including the following steps: S1. The crack images are labeled with bounding boxes. Local crack features are extracted through convolution, normalization, and activation functions. Multiple local crack features are then spatially pooled, normalized, and fused using the mean to form standard crack features. Specifically: Several clearly labeled images of road cracks from drones were selected as input images, and the crack targets in the input images were marked with bounding boxes. Obtain the target bounding box for each input image, and extract local features of the cracks through convolution, normalization, and activation functions; To avoid the instability of standard features caused by the influence of lighting, shadows, shooting angle or road texture on the features of a single image, instead of generating a standard template from a single image, spatial pooling, normalization and mean fusion of the local features of cracks from multiple crack images are performed to form a stable standard crack feature P. S2. Acquire images of the road to be detected, divide them into multiple image blocks of uniform size, and extract local road features block by block through convolution. Specifically: The images of the road to be detected are collected by drones. First, the images are scaled to 640×640 pixels and normalized as necessary to ensure that the size and pixel distribution of the input data meet the requirements of the subsequent network processing. Subsequently, the preprocessed road image to be detected is evenly divided into a plurality of image blocks according to a fixed size, for example, divided into 16×16 image blocks, and each image block corresponds to a local region in the road image to be detected; For each image block, road local features F are extracted block by block through convolution; S3. Acquiring road local features and standard crack features, calculating the similarity between the features through cosine similarity, and generating a similarity matrix according to the row and column positions of the road local feature image blocks in the road image to be detected, which specifically comprises: Performing cosine similarity calculation on the road local feature F and the standard crack feature P to measure the proximity between each image block and the standard crack feature, wherein the similarity calculation formula for the i-th image block is: wherein, is the similarity between the road local feature of the i-th image block and the standard crack feature, is a very small constant to prevent the denominator from being zero ( a larger value indicates that the region is more likely to contain cracks or road disease features; a smaller value indicates that the region is more likely to belong to background interference regions such as shadows, water stains, and pavement textures), represents the road local feature of the i-th image block, represents the standard crack feature; After completing the similarity calculation between the road local feature image blocks and the standard crack feature, the corresponding similarity values are mapped back to the block division grid according to the original spatial positions of the road local feature image blocks in the road image to be detected, so as to generate a similarity matrix consistent with the spatial layout of the image blocks; S4. Performing background suppression and crack feature enhancement on each image block region in the road image to be detected based on the similarity matrix, and outputting a background suppression enhanced map, which specifically comprises the following steps: (1) Calculating a background suppression weight, which specifically comprises: presetting a similarity threshold T, and when the similarity of a certain block is lower than the threshold, indicating that the region is more likely to belong to a background interference region; to avoid the problem that fine cracks are mistakenly suppressed due to the direct deletion of image regions by a hard threshold, the similarity is converted into a continuous background suppression weight: wherein, is the background suppression weight of the i-th block, is a Sigmoid function, T is the similarity threshold, k is a coefficient for adjusting the suppression intensity (when <T, When the value is small, the corresponding region is suppressed; when (S_i ≥ qT), Larger areas are preserved or enhanced. (2) The original image patch features are weighted and enhanced using background suppression weights to obtain the enhanced features after background suppression, specifically: Let the features of the original image patch be The enhancement features after background suppression are Then we have: in, This is an enhanced feature after background suppression. The original block features, For feature enhancement coefficients, The background suppression weights for the i-th block; When a certain region has a high similarity to the standard crack features, the background suppression weight is... Increase the enhancement features of the region after background suppression. Enhance; When the similarity of the region is low, the background suppression weight... Smaller, enhanced features after background suppression It will not be significantly enhanced, thus highlighting the crack features and suppressing background interference; Furthermore, the similarity threshold T is set to 0.4. When the similarity Si of the i-th image patch is greater than or equal to 0.4, it can be determined that the region has a high similarity to the standard crack feature, the background suppression weight wi is large, and the corresponding region is preserved or enhanced. When Si is less than 0.4, it can be determined that the region has a low similarity to the standard crack feature, the background suppression weight wi is small, and the corresponding region is suppressed. (3) Obtain the background suppression weight map, specifically: Background suppression weights for all image patches The background suppression weight map is obtained by restoring and interpolating the image to the size of the original image or feature map according to its spatial location. (4) The background suppression weight map is used to weight the image to be detected and output the background suppression enhancement map; S5. Input the background suppression enhancement map into the backbone feature extraction part of the YOLOv11 detection network, and perform convolutional normalization preprocessing, deep feature extraction, C2PSA attention enhancement (Cross-Channel Partial Self-Attention), and Spatial Pyramid Pooling-Fast processing in sequence to obtain multi-scale basic feature maps, including the following steps: S501, Convolutional Normalization Preprocessing: The input background suppression enhancement map first passes through the basic unit composed of Conv (convolutional layer), BN (batch normalization layer) and SiLU (activation function) to complete the preliminary feature mapping and normalization of the image, so as to improve the stability and convergence speed of subsequent network training. S502 and C3k2 modules (lightweight feature extraction modules): Feature extraction: The C3k2 module performs in-depth feature extraction and channel interaction on the preprocessed preliminary features to obtain extracted features; The YOLOv11 detection network backbone has multiple Conv and multiple C3k2 modules, and each Conv in the backbone network is followed by a C3k2 module. By using the residual structure and multi-branch convolution in the C3k2 module, the network's ability to represent different morphological features of cracks is enhanced while ensuring effective gradient propagation. S503, Spatial Pyramid Pooling (SPPF) expands the receptive field: The extracted features are input into the Spatial Pyramid Pooling (SPPF) module, which aggregates contextual information at different scales through continuous pooling operations, expands the network's receptive field, and obtains multi-scale contextual features that include local details and global semantic information. Furthermore, this step can enhance the ability of features to express crack structures at different scales without significantly increasing the computational load, enabling the network to retain local texture information of small cracks while also perceiving the overall structural features of longer cracks or complex disease areas. S504, C2PSA Attention Mechanism Enhancement: Input multi-scale contextual features into the C2PSA attention module to further enhance the features in terms of channel and spatial dimensions, obtain attention-enhanced features, and finally output multi-scale basic feature maps; The C2PSA attention module automatically learns and focuses on crack-related feature channels through a cross-channel local self-attention mechanism, while suppressing background noise channels such as shadows, water stains, and ordinary road surface textures, thereby highlighting key crack information and improving the model's attention to crack targets. S6. Input the multi-scale basic feature map into the feature fusion part of the YOLOv11 detection network. Use upsampling and Concat concatenation to fuse deep high-semantic features and shallow high-resolution detail features. Send the fused multi-scale features into the detection head to output the crack's location coordinates, size information, and confidence score, completing the final identification. This includes the following steps: S601, Multi-scale feature fusion: Upsample the multi-scale basic feature map to make its spatial size consistent with the shallow feature map of the previous layer, and obtain the upsampled high-level semantic features and the corresponding shallow high-resolution detail features. Subsequently, the upsampled high-level semantic features are concatenated with the corresponding shallow high-resolution detail features using the Concat channel concatenation method. The concatenated features are then input into the C3k2 module for feature integration and channel interaction to obtain the first fused feature. Furthermore, the first fusion feature is upsampled and concatted with the shallower high-resolution feature, and then fused again by the C3k2 module to obtain a high-resolution fusion feature for small-scale crack detection. After completing the top-down feature fusion, the high-resolution fused features are convolutionally downsampled to match their spatial size with the mesoscale fused features. Then, the downsampled features are concat-concatenated with the corresponding level features and fused through the C3k2 module to obtain fused features for mesoscale crack detection. We continue to perform convolutional downsampling on the mesoscale fusion features, concatenate them with the high-level semantic features, and then fuse them through the C3k2 module to obtain fusion features for large-scale crack detection. Finally, the three types of fused features—small-scale, medium-scale, and large-scale—are classified into a fused multi-scale feature, which is then input into the corresponding detection head for the localization and confidence prediction of crack targets at different scales. This method integrates semantic information from deep features without losing shallow detail information (such as crack edges and textures), generating a multi-scale fusion feature map that combines high resolution and strong semantic expression. S602, Detection Head Output: The fused multi-scale features are fed into the detection head. The detection head predicts the bounding box position, class confidence and target confidence of the target through parallel branches. Finally, it outputs the location coordinates (x, y, w, h), size information (width w and height h) and confidence score of the crack. Among them, the positioning coordinates are used to determine the location and range of the crack in the road image to be detected, the confidence score is used to characterize the reliability of the detection results, and serves as the basis for subsequent crack result screening, visualization annotation and road disease assessment. The final output crack detection results are used to determine the location, size and confidence level of the crack disease area in the road image to be detected, and can be further used for crack target box drawing, false detection screening, disease area statistics and road maintenance decision-making. Furthermore, after the detection head outputs candidate detection results, a parameter closed-loop optimization process is introduced. The system calculates the detection feedback loss for the t-th round based on the missed detections, false detections, and confidence deviations in the detection results, and uses this feedback loss to dynamically adjust the similarity threshold T and feature enhancement coefficient, thereby enabling the background suppression module to adaptively optimize according to different road scenes, lighting conditions, shadows, water stains, and complex texture interference.

[0020] Let the feedback loss of the t-th round of detection be: in, Let t be the feedback loss from the detection in round t. The loss from missed detection in round t (used to reflect the extent to which the actual crack area was not detected). The false detection loss for round t (used to reflect the degree to which background areas such as shadows, water stains, and ordinary road textures are misidentified as cracks). Let $t$ be the detection confidence bias loss (used to reflect the difference between the detection confidence and the expected confidence). , , These are the weighting coefficients; Based on the detection feedback results, the threshold update method can be expressed as: The enhancement coefficient update method can be expressed as: in, Let be the similarity threshold for round t+1. Update the step size for the parameters. For the threshold adjustment effect caused by false detection loss, The threshold reduction effect caused by the loss from missed detections, Let be the feature enhancement coefficients after the (t+1)th round update. Let be the feature enhancement coefficient for round t; In this way, the detection results output by the detection head are no longer just used as the final recognition result, but are further used as the feedback basis for updating the parameters of the background suppression module. This enables the algorithm to form a closed-loop optimization process of "similarity calculation - background suppression enhancement - multi-scale feature fusion - detection head output - detection feedback - parameter update", and finally output the updated and optimized crack detection results.

[0021] Example 2: Background suppression system based on UAV inspection, including: Crack feature construction module: Input crack images are labeled with target boxes, and local crack features are extracted through convolution, normalization and activation functions. Spatial pooling, normalization and mean fusion are performed on the local crack features of multiple crack images to obtain standard crack features. Road local feature extraction module: Input the road image to be detected, divide the image into multiple image blocks with uniform size, and extract the road local features corresponding to the image blocks through convolution feature extraction processing; Cosine similarity calculation module: acquires local road features and standard crack features, calculates the similarity between each road image block and the standard crack features through cosine similarity calculation, and generates a similarity matrix; Background suppression enhancement module: Input similarity matrix and local road features, calculate background suppression weights, perform weighted enhancement on the original image patch features based on the background suppression weights, obtain the enhanced features after background suppression, restore the size of the background suppression weights to obtain the background suppression weight map, perform weighted processing on the image to be detected based on the background suppression weight map, and output the background suppression enhancement map; YOLOv11 Detection Network: The background suppression enhancement map is input into the YOLOv11 detection network. Through convolutional normalization preprocessing, C3k2 deep feature extraction, C2PSA attention enhancement, and SPPF spatial pyramid pooling, a multi-scale basic feature map with both local detail information and global semantic information is obtained. The multi-scale basic feature map is input into the feature fusion part of the YOLOv11 detection network. Multi-scale fused features are generated through upsampling and Concat feature fusion. The detection head is then used to predict the location, size, and confidence of the crack target to obtain the crack detection result.

[0022] Experiment Example 1: Comparative Experiment on the Detection Performance of Road Defects The YOLOv11 detection network and the background suppression system based on UAV inspection of this invention were used for road defect detection scenarios and were deployed on UAVs. The detection target was road cracks. The UAV flight speed was 5 m / min. The input images were captured by a high-definition camera on the production line, preprocessed and adjusted to 640×640 pixels before being input into the model. The batch processing size was set to 40. Under the condition of P100 GPU and UAV-PDD2023 dataset, the performance of the two methods was detected (P - accuracy, R - recall, mAP50 - mean precision at an IoU threshold of 0.5, mAP50-95 - mean precision calculated under IoU thresholds of 0.5 to 0.95). The detection results are shown in Table 1.

[0023] Table 1 Comparison Results Table 1 shows that the YOLOv11 detection network has an accuracy of 0.84, a recall of 0.7, an mAP50 of 0.744, and an mAP50-95 of 0.453; while the present invention has an accuracy of 0.876, a recall of 0.729, an mAP50 of 0.798, and an mAP50-95 of 0.534. It can be seen that the present invention significantly improves the mAP50 and mAP50-95 for various road defects.

[0024] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such modifications or substitutions will 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 background suppression method based on unmanned aerial vehicle (UAV) inspection, characterized in that, Includes the following steps: S1. Mark the target bounding box on the crack image, extract the local features of the crack, and perform pooling and fusion processing on multiple local crack features to form standard crack features; S2. Acquire images of the road to be detected, divide them into multiple image blocks of uniform size, and extract local road features block by block through convolution. S3. Obtain local road features and standard crack features, calculate the similarity between features using cosine similarity, and generate a similarity matrix based on the row and column positions of the local road feature image blocks in the road image to be detected. S4. Based on the similarity matrix, perform background suppression and crack feature enhancement on each image block region in the road image to be detected, and output a background suppression and enhancement map. This includes the following steps: (1) Calculate the background suppression weights; (2) Use background suppression weights to weight and enhance the original image patch features to obtain the enhanced features after background suppression; (3) Obtain the background suppression weight map; (4) The background suppression weight map is used to weight the image to be detected and output the background suppression enhancement map; S5. Input the background suppression enhancement map into the backbone feature extraction part of the YOLOv11 detection network, and perform convolutional normalization preprocessing, deep feature extraction, C2PSA attention enhancement and spatial pyramid pooling SPPF processing in sequence to obtain multi-scale basic feature maps. S6. Input the multi-scale basic feature map into the feature fusion part of the YOLOv11 detection network. Use upsampling and stitching to fuse deep high semantic features and shallow high resolution detail features to obtain fused multi-scale features. Then, send the fused multi-scale features into the detection head to output the crack's location coordinates, size information and confidence level to complete the final identification.

2. The background suppression method based on UAV inspection according to claim 1, characterized in that, Step S1 specifically involves: Multiple clearly labeled images of road cracks from drones were selected as input images, and the crack targets in the input images were labeled with bounding boxes. Obtain the bounding box annotations of the input image, and extract local features of the cracks through convolution, normalization, and activation functions; Spatial pooling, normalization, and mean fusion are performed on the local crack features of multiple crack images to form stable standard crack features.

3. The background suppression method based on UAV inspection according to claim 1, characterized in that, Step S2 specifically involves: Images of the road to be inspected are collected by drones. First, the images are sized uniformly and then normalized. The preprocessed road image to be detected is evenly divided into multiple image blocks of a fixed size; For each image block, local road features are extracted block by block through convolution.

4. The background suppression method based on UAV inspection according to claim 1, characterized in that, Step S3 specifically involves: Cosine similarity is calculated between local road features and standard crack features to measure the closeness between each image patch and the standard crack features. Based on the original spatial location of each road local feature image block in the image of the road to be detected, the corresponding similarity value is mapped back to the block grid to generate a similarity matrix that is consistent with the spatial layout of the image blocks.

5. The background suppression method based on UAV inspection according to claim 1, characterized in that, In step (1), the background suppression weight is calculated using the following formula: ; in, The background suppression weights for the i-th block are... Here, is the Sigmoid function, T is the similarity threshold, and k is a coefficient that adjusts the suppression strength.

6. The background suppression method based on UAV inspection according to claim 5, characterized in that, In step (2), the features of the original image patch are assumed to be... The formula for calculating the enhanced features after background suppression is: ; in, This is an enhanced feature after background suppression. The original block features, For feature enhancement coefficients, is the background suppression weight for the i-th block.

7. The background suppression method based on UAV inspection according to claim 1, characterized in that, Step S5 specifically includes the following steps: S501. The input background suppression enhancement map first passes through basic units consisting of convolutional layers, batch normalization layers, and SiLU activation functions to form the initial feature mapping and standardization of the image. S502. The C3k2 module performs in-depth feature extraction and channel interaction processing on the preliminary features of the executed image to obtain the extracted features. S503. Input the extracted features into the Spatial Pyramid Pooling (SPPF) module, and aggregate contextual information at different scales through continuous pooling operations to obtain multi-scale contextual features. S504. Input the multi-scale contextual features into the C2PSA attention module to enhance the features in both channel and spatial dimensions, obtain attention-enhanced features, and finally output the multi-scale basic feature map.

8. The background suppression method based on UAV inspection according to claim 1, characterized in that, Step S6 includes the following steps: S601. Upsample the multi-scale basic feature map step by step and merge it with the corresponding shallow features; then merge it with the high-level features through step-by-step downsampling. After integration by the C3k2 module, the fused multi-scale features are obtained. S602. The fused multi-scale features are fed into the detection head. The detection head predicts the bounding box position, class confidence and target confidence of the target through parallel branches, and outputs the location coordinates, size information and confidence score of the crack.

9. A background suppression system based on unmanned aerial vehicle (UAV) inspection, characterized in that, The system is used in the background suppression method based on UAV inspection as described in any one of claims 1-8, and the system comprises: Crack feature construction module: Input crack images, mark target boxes, extract local crack features, and perform pooling and fusion processing on the local crack features of multiple crack images to obtain standard crack features; Road local feature extraction module: Input the road image to be detected, divide the image into multiple image blocks with uniform size, and extract the road local features corresponding to the image blocks through convolution feature extraction processing; Cosine similarity calculation module: acquires local road features and standard crack features, calculates the similarity between each road image block and the standard crack features through cosine similarity calculation, and generates a similarity matrix; Background suppression enhancement module: Input similarity matrix and local road features, calculate background suppression weights, perform weighted enhancement on the original image patch features based on the background suppression weights, obtain the enhanced features after background suppression, restore the size of the background suppression weights to obtain the background suppression weight map, perform weighted processing on the image to be detected based on the background suppression weight map, and output the background suppression enhancement map; YOLOv11 Detection Network: The background suppression enhancement map is input into the YOLOv11 detection network. Through convolutional normalization preprocessing, C3k2 deep feature extraction, C2PSA attention enhancement, and SPPF spatial pyramid pooling, a multi-scale basic feature map with both local detail information and global semantic information is obtained. The multi-scale basic feature map is input into the feature fusion part of the YOLOv11 detection network. Multi-scale fused features are generated through upsampling and Concat feature fusion. The detection head is then used to predict the location, size, and confidence of the crack target to obtain the crack detection result.

Images