Road defect detection method, system and medium
By improving the feature interaction enhancement fusion module and dynamic detection head of the YOLOv8 model, the detection challenges of YOLOv8 in multi-scale and complex scenarios have been solved, achieving efficient and accurate road defect detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUZHOU UNIVERSITY
- Filing Date
- 2026-01-16
- Publication Date
- 2026-04-21
AI Technical Summary
Existing mainstream detection models such as YOLOv8 are unable to effectively capture multi-scale features in road defect detection, and features are easily submerged by noise in complex scenarios, resulting in high false detection and false negative rates, making it difficult to accurately model diverse types of defects.
The original YOLOv8 model's channel stitching module is replaced by a Feature Interaction Enhancement Fusion Module (FIEF), and a dynamic detection head (DyHead) is introduced to enhance multi-scale feature extraction and adaptive detection capabilities.
It significantly improves the model's ability to extract features from multi-scale and low-contrast defects, enhances detection accuracy and robustness, while maintaining computational efficiency and adapting to different detection tasks.
Smart Images
Figure CN121904472A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and road inspection technology, and in particular to a method, system and medium for detecting road defects. Background Technology
[0002] With the rapid development of intelligent transportation systems and digital management of road infrastructure, automatic detection technology for road surface defects plays an increasingly important role in improving inspection efficiency and ensuring driving safety. However, in actual road inspection scenarios, defect detection still faces severe challenges: road surface defects vary significantly in scale, ranging from micro-cracks to large potholes, making it difficult for models to effectively capture multi-scale features simultaneously; under complex road backgrounds, shadow occlusion, or uneven lighting conditions, the contrast between defects and the background is low, and their features are easily drowned out by noise; in addition, defects of the same type have diverse forms, while defects of different types sometimes have similar visual features. These factors together lead to a high rate of false positives and false negatives.
[0003] While widely adopted mainstream detection models such as YOLOv8 achieve a good balance between speed and accuracy, their native architecture still has significant limitations when dealing with the aforementioned complex scenarios. Specifically, their neck network feature fusion mechanism is relatively simple, failing to fully realize cross-level and cross-scale deep semantic information interaction and enhancement. Furthermore, their fixed detection head design lacks dynamic adaptability to multiple tasks such as classification and localization, making it difficult to accurately model and adjust for the diverse forms of road defects and their complex contextual relationships. Summary of the Invention
[0004] The purpose of this invention is to provide a road defect detection method that enhances the performance of multi-scale defect feature extraction and complex scene detection.
[0005] To address the aforementioned technical problems, embodiments of the present invention provide a road defect detection method, comprising the following steps: Acquire road surface images; The road surface image is input into a pre-trained road defect detection model, which outputs the type and location information of defects present in the road surface image. The road defect detection model is an improved model based on YOLOv8. The road defect detection model replaces all channel stitching modules in the Neck part of the original YOLOv8 model with the Feature Interaction Enhancement Fusion (FIEF) module, and replaces the detection head with the dynamic detection head DyHead. The FIEF is based on the Shallow Detail Fusion (SDFM) module, and adds a channel attention mechanism (SE) after the second Sigmoid operation of the original SDFM module. The Backbone part of the road defect detection model extracts features from the road surface image to obtain multi-scale features. The FIEF part performs bidirectional cross-layer guidance and dual attention weighted fusion of the multi-scale features and the shallow detail features passed through upsampling in the neck network. The channel attention mechanism SE is used to recalibrate the channels of the fused features and output the enhanced multi-scale feature representation of the image. The dynamic detection head DyHead performs adaptive weighting and aggregation of the enhanced multi-scale features output from the C2f module in the Neck part based on scale perception, spatial perception and task interaction, and outputs the defect category and location information in the road surface image.
[0006] In some optional embodiments, the Feature Interaction Enhancement Fusion (FIEF) module specifically follows the following process: The first input feature and the second input feature are fused along the channel dimension to obtain the fused feature; global average pooling, CBS module processing, CB module processing, and Sigmoid activation are sequentially performed on the fused feature to generate the first fused weight; the first fused weight is multiplied element-wise with the fused feature, and then added element-wise with the fused feature to obtain the primary enhanced feature; the CBS module is a combination module consisting of a 1x1 convolutional layer, a BatchNorm layer, and a SiLU activation function connected in sequence, and the CB module is a combination module consisting of a 1x1 convolutional layer and a BatchNorm layer connected in sequence; The primary enhancement features are input into the CBS module for processing to obtain weighted base features. These weighted base features are then simultaneously input into the spatial attention path and the second channel attention path to generate fine-grained spatial fusion weights and fine-grained channel fusion weights, respectively. The fine-grained spatial fusion weights and fine-grained channel fusion weights are multiplied element-wise and then activated by a Sigmoid function to generate fine-grained fusion weights. Simultaneously, the primary enhancement features are split into a first channel part and a second channel part. The spatial attention path is a combined module consisting of the CBS module and the CB module connected sequentially, and the second channel attention path is a combined module consisting of the GAP module and two CBS modules connected sequentially. The fine-grained fusion weights are multiplied element-wise with the first channel component to obtain the first weighted feature; the complementary weights of the fine-grained fusion weights are multiplied element-wise with the second channel component to obtain the second weighted feature; the first weighted feature is input into the SE module for channel recalibration, and then added element-wise with the second input feature to obtain the first fusion result; the second weighted feature is added element-wise with the first input feature to obtain the second fusion result; the first fusion result and the second fusion result are concatenated along the channel dimension to output the multi-scale fusion feature.
[0007] In some optional embodiments, the dynamic detection head DyHead specifically follows the following process: It receives the enhanced multi-scale features output from the C2f module in the Neck section, adjusts the multi-scale features to a uniform scale, and forms a three-dimensional feature tensor containing hierarchical, spatial, and channel dimensions. A 3D feature tensor is input into a dynamic feature enhancement block, which sequentially performs scale-aware attention, spatial-aware attention, and task-aware attention operations. The scale-aware attention operation applies attention to the hierarchical dimensions of the 3D feature tensor to adaptively fuse different feature levels associated with the target scale. The spatial-aware attention operation applies attention to the spatial dimensions of the 3D feature tensor and performs adaptive sparse sampling through deformable convolution to focus on discriminative spatial regions and integrate cross-level spatial features. The task-aware attention operation applies attention to the channel dimensions of the 3D feature tensor and adjusts the activation state of each feature channel through a dynamic learning mechanism to adapt it to different downstream detection tasks. The dynamic feature enhancement blocks can be sequentially stacked and repeatedly executed to progressively enhance the input features. The enhanced features processed by the dynamic feature enhancement block are output, and target classification and localization are performed based on the enhanced features.
[0008] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the road defect detection method described above.
[0009] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when run by a processor, is capable of executing the road defect detection method described above.
[0010] The road defect detection method provided by this invention has at least the following beneficial effects: This invention significantly enhances the model's feature extraction and fusion capabilities for multi-scale and low-contrast defects by replacing the channel stitching operation of the original YOLOv8 model with a Feature Interaction Enhancement Fusion Module (FIEF). Simultaneously, by introducing a dynamic detection head (DyHead), the model can adaptively optimize classification and localization tasks, improving the localization accuracy and scene robustness for irregular defects. This solution achieves a significant improvement in detection accuracy while maintaining computational efficiency comparable to the original model, demonstrating good engineering practicality and providing an efficient and reliable technical solution for automated road inspection. Attached Figure Description
[0011] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0012] Figure 1 This is a flowchart of a road defect detection method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a road defect detection model provided according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a Feature Interaction Enhancement Fusion Module (FIEF) provided according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a dynamic detection head DyHead provided according to an embodiment of the present invention; Figure 5 This is an example diagram of road defect categories provided according to an embodiment of the present invention. Figure 5 (a) is a longitudinal crack. Figure 5 (b) is a transverse crack. Figure 5 (c) is a network of cracks. Figure 5 (d) represents a pit or groove. Figure 5 (e) is for repair; Figure 6 This is a schematic diagram illustrating the detection results of various defect types in the China_Drone dataset using different methods provided by an embodiment of the present invention. (a) shows a longitudinal crack. Figure 6 (b) is a transverse crack. Figure 6 (c) is a network of cracks. Figure 6 (d) represents a pit or groove. Figure 6 (e) is for repair; Figure 7 This is a schematic diagram illustrating the detection results of various defect types in the China_Motorbike dataset using different methods provided by an embodiment of the present invention. Figure 7 (a) is a longitudinal crack. Figure 7 (b) is a transverse crack. Figure 7 (c) is a network of cracks. Figure 7 (d) represents a pit or groove. Figure 7 (e) is for repair; Figure 8 This is a schematic diagram illustrating the detection results of various defect types in the RDD_China dataset using different methods provided by an embodiment of the present invention. Figure 8 (a) is a longitudinal crack. Figure 8 (b) is a transverse crack. Figure 8 (c) is a network of cracks. Figure 8 (d) represents a pit or groove. Figure 8(e) is for repair. Detailed Implementation
[0013] 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 in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0014] One embodiment of the present invention relates to a road defect detection method. The implementation details of the road defect detection method of this embodiment are described in detail below. The following content is only for the convenience of understanding and is not necessary for implementing this solution.
[0015] The specific process of the road defect detection method in this embodiment can be as follows: Figure 1 As shown, it includes: Step 101: Obtain a road surface image; Step 102: Input the road surface image into the pre-trained road defect detection model and output the defect category and location information in the road surface image; wherein, the road defect detection model is an improved model based on YOLOv8; the road defect detection model replaces all channel stitching modules in the Neck part of the original YOLOv8 model with the Feature Interaction Enhancement Fusion (FIEF) module, and replaces the detection head with the dynamic detection head DyHead; the FIEF is based on the Shallow Detail Fusion (SDFM) module, and adds a channel attention mechanism SE after the second Sigmoid operation of the original SDFM module; The Backbone part of the road defect detection model extracts features from the road surface image to obtain multi-scale features. The FIEF part performs bidirectional cross-layer guidance and dual attention weighted fusion of the multi-scale features and the shallow detail features passed through upsampling in the neck network. The channel attention mechanism SE is used to recalibrate the channels of the fused features and output the enhanced multi-scale feature representation of the image. The dynamic detection head DyHead performs adaptive weighting and aggregation of the enhanced multi-scale features output from the C2f module in the Neck part based on scale perception, spatial perception and task interaction, and outputs the defect category and location information in the road surface image.
[0016] Road defect detection model such as Figure 2 As shown in the figure, the channel stitching module of the original YOLOv8 model's Neck part is replaced with the Feature Interaction Enhancement Fusion Module (FIEF), and the detection head of the original YOLOv8 model is replaced with the dynamic detection head DyHead.
[0017] Feature Interaction Enhancement Fusion Module (FIEF) Figure 3 As shown in the figure, the first input feature and the second input feature are fused along the channel dimension to obtain the fused feature. Global average pooling, CBS module processing, CB module processing, and Sigmoid activation are sequentially performed on the fused feature to generate the first fused weight. The first fused weight is then multiplied element-wise with the fused feature, and then added element-wise to the fused feature to obtain the primary enhancement feature. The CBS module is a combination module consisting of a 1x1 convolutional layer, a BatchNorm layer, and a SiLU activation function connected sequentially. The CB module is a combination module consisting of a 1x1 convolutional layer and a BatchNorm layer connected sequentially. The primary enhancement feature is input into the CBS module for processing to obtain the weight-generated base feature. The weight-generated base feature is simultaneously input into the spatial attention path and the second channel attention path to generate fine-grained spatial fusion weights and fine-grained channel fusion weights, respectively. The fine-grained spatial fusion weights and fine-grained channel fusion weights are then multiplied element-wise with the first fused weight and the second channel attention path. After element-wise multiplication of the channel fusion weights, a fine-grained fusion weight is generated by sigmoid activation. Simultaneously, the primary enhancement feature is split into a first channel component and a second channel component. The spatial attention path is a combined module consisting of a CBS module and a CB module connected sequentially, and the second channel attention path is a combined module consisting of a GAP module and two CBS modules connected sequentially. The fine-grained fusion weight is element-wise multiplied with the first channel component to obtain the first weighted feature. The complementary weights of the fine-grained fusion weight are element-wise multiplied with the second channel component to obtain the second weighted feature. The first weighted feature is input into the SE module for channel recalibration and then element-wise added to the second input feature to obtain the first fusion result. The second weighted feature is element-wise added to the first input feature to obtain the second fusion result. The first and second fusion results are concatenated along the channel dimension to output a multi-scale fusion feature.
[0018] Dynamic detection head DyHead such as Figure 4As shown in the diagram, the enhanced multi-scale features received from the C2f module in the Neck section are adjusted to a uniform scale to form a three-dimensional feature tensor containing hierarchical, spatial, and channel dimensions. This three-dimensional feature tensor is then input to a dynamic feature enhancement block, which sequentially performs scale-aware attention, spatial-aware attention, and task-aware attention operations. The scale-aware attention operation applies attention to the hierarchical dimension of the three-dimensional feature tensor to adaptively fuse different feature levels associated with the target scale. The spatial-aware attention operation applies attention to the spatial dimension of the three-dimensional feature tensor, performing adaptive sparse sampling through deformable convolution to focus on discriminative spatial regions and integrate cross-level spatial features. The task-aware attention operation applies attention to the channel dimension of the three-dimensional feature tensor, adjusting the activation state of each feature channel through a dynamic learning mechanism to adapt it to different downstream detection tasks. The dynamic feature enhancement blocks can be stacked sequentially and executed repeatedly to progressively enhance the input features. The enhanced features processed by the dynamic feature enhancement blocks are output, and target classification and localization are performed based on these enhanced features.
[0019] A specific example: Dataset introduction and preparation: This embodiment uses the RDD2022 (Road Damage Dataset 2022) dataset for model training and validation. This dataset was collected from six countries: China, Japan, Czech Republic, Norway, the United States, and India, and contains a total of 47,420 high-resolution road surface images and corresponding annotations. For the Chinese region data (4,373 annotated images), it is further divided into two subsets with different imaging characteristics:
[0020] China_Motorbike: The images were captured by a smartphone mounted on the motorcycle. The angled shooting and high resolution closely resemble the perspective of a human inspector, which is beneficial for capturing local details.
[0021] China_Drone: The images were captured by drones from above, offering a wide field of view, relatively low resolution, and greater scene coverage.
[0022] The two subsets were merged into the RDD_China dataset to provide richer image features and defect types. The distribution of defect data within the dataset is shown in Table 1.
[0023] Table 1. Distribution of Defect Data within the Dataset Example diagram of road defect categories Figure 5 As shown, Figure 5 (b) is a transverse crack. Figure 5 (c) is a network of cracks. Figure 5 (d) represents a pit or groove. Figure 5 (e) For repair, the data labels in the figure cover five typical road defects: longitudinal cracks (D00), transverse cracks (D10), network cracks (D20), potholes (D30), and repair areas (D40). All datasets are divided into training, validation, and test sets in a 7:1:2 ratio.
[0024] Experimental environment and hyperparameter settings: The experiment is based on the PyTorch deep learning framework, using YOLOv8 as the baseline model. The experimental environment configuration is as follows:
[0025] Hardware: Ubuntu 22.04.3 LTS operating system, Intel Core i9-12900k processor, NVIDIA GeForce RTX 3080 Ti graphics card (12GB VRAM), 32GB RAM.
[0026] Software: Python 3.8.13, PyTorch 1.13.0, CUDA 11.7.
[0027] Training hyperparameters: Input image resolution 640×640, batch size 8, Adam optimizer used, initial learning rate 0.01, weight decay 0.0005, momentum 0.937, training for 150 epochs. The learning rate was reduced to 0.1 times its original value at 67% and 89% of the total epochs. Data augmentation only used random horizontal flipping.
[0028] Ablation experiment analysis: To verify the effectiveness of the proposed improved module, systematic ablation experiments were conducted on three datasets. The ablation experiment results on the China_Drone dataset are shown in Table 2, the performance results on the China_Motorbike dataset are shown in Table 3, and the performance results on the RDD_China dataset are shown in Table 4.
[0029] Table 2 Ablation experiment results on the China_Drone dataset Table 3 Performance results on the China_Motorbike dataset Table 4. Performance results on the RDD_China dataset. On the China_Drone dataset (low resolution, top-down view), using the Feature Interaction Enhancement Fusion Module (FIEF) alone improved mAP@0.5 from the baseline of 62.4% to 63.5%; using the Dynamic Detection Head (DyHead) alone improved it to 64.7%; and using both together achieved 67.5%, demonstrating the synergistic effect between the modules.
[0030] On the China_Motorbike dataset (high resolution, tilted view), the combined use of FIEF and DyHead achieved the best results, with mAP@0.5 reaching 93.7%, an improvement of 2.3 percentage points over the baseline (91.4%), demonstrating the model's advantage in processing clear image details.
[0031] On the comprehensive RDD_China dataset, the combined use of FIEF and DyHead improved mAP@0.5 to 83.5%, an improvement of 4.1 percentage points from the baseline (79.4%), validating the stability and generalization ability of the method in diverse scenarios.
[0032] Experiments show that the FIEF module enhances the model's ability to capture complex features by strengthening the dependencies between features; DyHead improves the recognition accuracy of multi-scale targets by adaptively adjusting the detection strategy.
[0033] Comparative experiments and performance evaluation: The improved model proposed in this embodiment was comprehensively compared with current mainstream object detection algorithms on three datasets, including YOLOv8, YOLOv9, YOLOv10, YOLOv11, YOLOv12, Faster R-CNN, Mask R-CNN (ConvNeXt), LiteYOLO-ID, MAF-YOLOv2-n, and MAF-YOLOv2-s. Table 5 shows the comparison of results for different algorithms on the China_Drone dataset, Table 6 shows the comparison of results for different algorithms on the China_Motorbike dataset, and Table 7 shows the comparison of results for different algorithms on the RDD_China dataset.
[0034] Table 5 Comparison of results of different algorithms on the China_Drone dataset Table 6. Comparison of results of different algorithms on the China_Motorbike dataset Table 7 Comparison of results of different algorithms on the RDD_China dataset On the China_Drone dataset, our method outperforms the listed comparison models in mAP@0.5 (67.5%), mAP@0.5:0.95 (40.7%), and accuracy (72.1%), while maintaining a reasonable computational cost (11.4 GFLOPs).
[0035] On the China_Motorbike dataset, this method achieved the highest mAP@0.5 (93.7%), mAP@0.5:0.95 (65.0%), precision (91.1%), and recall (89.5%), significantly outperforming other models.
[0036] On the RDD_China dataset, this method also achieved the best results with 83.5% mAP@0.5 and 54.1% mAP@0.5:0.95, demonstrating its comprehensive performance advantage.
[0037] Comparative experiments show that our method significantly surpasses existing methods in detection accuracy, while achieving a good balance between model complexity and computational efficiency.
[0038] Performance verification of defect detection to ensure traffic safety: In-depth analysis was conducted on defect categories that have a significant impact on traffic safety (such as longitudinal cracks and potholes). The mAP50 (%) of each defect category in the China_Drone dataset is shown in Table 8, the mAP50 (%) of each defect category in the China_Motorbike dataset is shown in Table 9, and the mAP50 (%) of each defect category in the RDD_China dataset is shown in Table 10.
[0039] Table 8. mAP50 (%) for each defect category in the China_Drone dataset Table 9. mAP50 (%) for each defect category in the China_Motorbike dataset Table 10. mAP50 (%) for each defect category in the RDD_China dataset In longitudinal crack detection, our method achieves 83.2% accuracy on the RDD_China dataset, outperforming most comparative models.
[0040] In pit detection, our method achieved an mAP of 84.3% on the RDD_China dataset, significantly outperforming the comparison model and demonstrating its accurate identification capability for high-risk 3D defects.
[0041] This indicates that the method can effectively identify key road surface defects that pose a direct threat to driving safety and has practical application value.
[0042] Visual analysis: Detection results of different methods for various defect types in the China_Drone dataset, as follows: Figure 6 As shown, Figure 6 (a) is a longitudinal crack. Figure 6 (b) is a transverse crack. Figure 6 (c) is a network of cracks. Figure 6 (d) represents a pit or groove. Figure 6 (e) To illustrate the repair, the figure compares the detection results of the proposed method and the base model YOLOv8 on the China_Drone dataset. It can be seen that the detection results of the proposed method are highly consistent with the true labels, with only one false positive and no false negatives; while YOLOv8 exhibits multiple false positives and two obvious false negatives. Experimental results show that the proposed method significantly outperforms the benchmark model in terms of detection accuracy and robustness.
[0043] Detection results of different methods for various defect types in the China_Motorbike dataset, as follows: Figure 7 As shown, Figure 7 (a) is a longitudinal crack. Figure 7 (b) is a transverse crack. Figure 7 (c) is a network of cracks. Figure 7 (d) represents a pit or groove. Figure 7 (e) For the repair, the figure compares the detection results of the proposed method and the base model YOLOv8 on the China_Motorbike dataset. It can be seen that the detection results of the present invention highly match the true labels, with no false positives or false negatives; while YOLOv8 has multiple false positives and significantly lower detection accuracy. Experimental results show that the present invention can more reliably identify road defects in complex scenarios and has stronger practical application capabilities.
[0044] Detection results of different methods for various defect types in the RDD_China dataset, as follows: Figure 8 As shown, Figure 8 (a) is a longitudinal crack. Figure 8 (b) is a transverse crack. Figure 8 (c) is a network of cracks. Figure 8 (d) represents a pit or groove. Figure 8 (e) To illustrate the repair, the figure compares the detection results of the proposed method and the base model YOLOv8 on the RDD_China dataset. It can be seen that the detection results of the present invention are highly consistent with the true labels, with no false positives or false negatives; while YOLOv8 has multiple false positives and two false negatives, resulting in significantly lower detection accuracy.
[0045] The detection results of this method closely match the ground truth (GT) annotations, with significantly fewer false positives and false negatives than YOLOv8. For example, in images from the China_Motorbike and RDD_China datasets with complex backgrounds, YOLOv8 misclassified shadows and textures as defects in several instances, while this method accurately focused on the true defect areas. Visualization results further confirm that this method has stronger robustness and higher detection reliability in complex scenes.
[0046] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.
[0047] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.
[0048] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor 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.
[0049] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.
Claims
1. A method for detecting road defects, characterized in that, The method includes: Acquire road surface images; The road surface image is input into a pre-trained road defect detection model, which outputs the type and location information of defects present in the road surface image. The road defect detection model is an improved model based on YOLOv8. The road defect detection model replaces all channel stitching modules in the Neck part of the original YOLOv8 model with the Feature Interaction Enhancement Fusion (FIEF) module, and replaces the detection head with the dynamic detection head DyHead. The FIEF is based on the Shallow Details Fusion (SDFM) module, and adds a channel attention mechanism (SE) after the second Sigmoid operation of the original SDFM module. The Backbone part of the road defect detection model extracts features from the road surface image to obtain multi-scale features. The FIEF part performs bidirectional cross-layer guidance and dual attention weighted fusion of the multi-scale features and the shallow detail features passed through upsampling in the neck network. The channel attention mechanism SE is used to recalibrate the channels of the fused features and output the enhanced multi-scale feature representation of the image. The dynamic detection head DyHead performs adaptive weighting and aggregation of the enhanced multi-scale features output from the C2f module in the Neck part based on scale perception, spatial perception and task interaction, and outputs the defect category and location information in the road surface image.
2. The road defect detection method as described in claim 1, characterized in that, The Feature Interaction Enhancement Fusion Module (FIEF) operates as follows: The first input feature and the second input feature are fused along the channel dimension to obtain the fused feature; Global average pooling, CBS module processing, CB module processing, and Sigmoid activation are sequentially applied to the fused features to generate the first fused weights. The first fused weights are then multiplied element-wise with the fused features, and then added element-wise with the fused features to obtain the primary enhanced features. The CBS module is a combination module consisting of a 1x1 convolutional layer, a BatchNorm layer, and a SiLU activation function connected in sequence. The CB module is a combination module consisting of a 1x1 convolutional layer and a BatchNorm layer connected in sequence. The primary enhancement features are input into the CBS module for processing to obtain weighted basis features; The weights are generated by simultaneously inputting the basis features into the spatial attention path and the second channel attention path to generate fine-grained spatial fusion weights and fine-grained channel fusion weights, respectively. After multiplying the fine-grained spatial fusion weights and the fine-grained channel fusion weights element-wise, the fine-grained fusion weights are generated by sigmoid activation; at the same time, the primary enhancement features are split into a first channel part and a second channel part; the spatial attention path is a combined module composed of CBS module and CB module connected in sequence, and the second channel attention path is a combined module composed of GAP module and two CBS modules connected in sequence. The fine-grained fusion weights are multiplied element-wise with the first channel component to obtain the first weighted feature; the complementary weights of the fine-grained fusion weights are multiplied element-wise with the second channel component to obtain the second weighted feature; the first weighted feature is input into the SE module for channel recalibration, and then added element-wise with the second input feature to obtain the first fusion result; the second weighted feature is added element-wise with the first input feature to obtain the second fusion result; the first fusion result and the second fusion result are concatenated along the channel dimension to output the multi-scale fusion feature.
3. The road defect detection method as described in claim 1, characterized in that, The specific process of the dynamic detection head DyHead is as follows: It receives the enhanced multi-scale features output from the C2f module in the Neck section, adjusts the multi-scale features to a uniform scale, and forms a three-dimensional feature tensor containing hierarchical, spatial, and channel dimensions. The three-dimensional feature tensor is input into the dynamic feature enhancement block, which sequentially performs scale-aware attention operation, spatial awareness attention operation and task-aware attention operation. The scale-aware attention operation applies attention to the hierarchical dimensions of the three-dimensional feature tensor to adaptively fuse different feature levels associated with the target scale. The spatial awareness attention operation applies attention to the spatial dimension of the three-dimensional feature tensor and performs adaptive sparse sampling through deformable convolution to focus on discriminative spatial regions and integrate cross-level spatial features. The task-aware attention operation applies attention to the channel dimension of the three-dimensional feature tensor and regulates the activation state of each feature channel through a dynamic learning mechanism to adapt it to different downstream detection tasks; the dynamic feature enhancement blocks can be stacked sequentially and executed repeatedly to progressively enhance the input features. The enhanced features processed by the dynamic feature enhancement block are output, and target classification and localization are performed based on the enhanced features.
4. A computer system, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the road defect detection method as described in any one of claims 1 to 3.
5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, is capable of performing the road defect detection method as defined in any one of claims 1 to 3.