Improved YOLO11 Lightweight Detection Method and System for Road Dysfunctions for Routine Inspections

Abstract

This invention discloses an improved lightweight YOLO11 road surface defect detection method and system for routine inspections. The method includes: acquiring and preprocessing road surface inspection images; inputting input feature tensors into a backbone network to extract multi-scale feature maps, with the backbone network using EfficientNetV2-S and shallow layers replaced by Fused-MBConv instead of MBConv; inputting the multi-scale feature maps into a SlimNeck network for cross-scale fusion, with the neck downsampling path configured with a GSConv module, and the features concatenated and then introduced into a VoVGSCSP module; inputting the fused features into a detection head to output candidate bounding boxes and confidence scores, with the WIoUv3 loss function used during the training phase to optimize bounding box regression; and outputting the defect category, confidence score, and coordinates after non-maximum suppression. This invention significantly reduces the number of parameters and computational load through a lightweight backbone and efficient neck design, achieving real-time detection while maintaining high detection accuracy. It is suitable for routine road surface inspection scenarios with limited computing power, such as vehicle-mounted and edge-based systems.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent transportation technology, specifically to an improved lightweight detection method and system for YOLO11 road surface defects for routine inspections. Background Technology

[0002] With the increasing demand for routine inspections of highways, urban roads, and industrial park roads, the ability to quickly identify road surface defects such as cracks, potholes, and repair cracks using vehicle-mounted, edge-mounted, or mobile devices has become a key issue in intelligent road maintenance. Traditional manual inspection methods suffer from drawbacks such as low efficiency, strong subjectivity, high risk, and weak data logging capabilities. While automated detection solutions using hardware such as LiDAR and specialized linear array equipment offer higher accuracy, their high equipment investment and maintenance costs make them difficult to promote in large-scale, routine scenarios.

[0003] In recent years, target detection methods based on convolutional neural networks, especially the YOLO series, have become the mainstream technical approach for pavement distress identification due to their advantages such as end-to-end detection, flexible deployment, and fast processing speed. Existing solutions can be roughly divided into two categories: First, directly using general target detection models for distress identification. Although this type of solution has a certain level of accuracy, it has a large number of model parameters and high computational complexity, making it difficult to maintain real-time detection on edge devices. Second, using general lightweight backbone networks for compression. Although this type of solution can reduce the amount of computation, it often suffers from missed detections, false detections, or inaccurate localization when dealing with thin cracks, network cracks, and irregularly shaped distresses due to insufficient shallow texture extraction and multi-scale fusion capabilities.

[0004] Existing technologies suffer from at least the following shortcomings: First, it is difficult to balance accuracy and efficiency. While complex structures can enhance feature representation, they significantly increase computational load, hindering deployment. Second, they lack the ability to regress boundaries for difficult samples. Cracks and other defects are elongated, discontinuous, and have blurred boundaries. When using ordinary IoU loss, the model often struggles to effectively focus on samples of moderate quality that require the most optimization. Third, their ability to fuse multi-scale defects is limited. Insufficient integration between shallow texture information and deep semantic information can easily lead to missed detections of small cracks and incomplete boundaries for large-area defects. Fourth, they lack engineering deployment friendliness, failing to meet the comprehensive requirements of high frame rate, low latency, and low computational cost in vehicle-mounted inspection scenarios.

[0005] Therefore, how to design a lightweight detection method that can significantly reduce the number of model parameters and computational complexity while maintaining high detection accuracy, and enhance the ability to locate multi-scale and irregular road surface defects, making it suitable for routine and mobile inspections, has become an urgent technical problem to be solved in this field. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing pavement distress detection methods in terms of lightweight design, positioning accuracy, and engineering deployment. It proposes an improved YOLO11 lightweight pavement distress detection method and system for routine inspections. Through lightweight reconstruction of the backbone network, efficient feature fusion of the neck network, and optimization of bounding box regression loss, it achieves rapid and accurate detection of pavement distresses such as transverse cracks, longitudinal cracks, mesh cracks, repair cracks, and potholes.

[0007] To achieve the above objectives, this invention focuses on solving the following technical problems: First, significantly reducing the number of model parameters and computational load without substantially reducing detection accuracy, enabling its deployment on embedded devices, vehicle terminals, or mobile devices; Second, enhancing the model's ability to extract features and fuse multiple scales for minor cracks, irregularly shaped cracks, and pavement defects with significant scale differences; Third, by optimizing the bounding box regression loss function, the model can allocate more optimization resources to samples of moderate quality but difficult to locate during the training phase, thereby improving the localization accuracy of complex defect targets; Fourth, establishing a replicable, implementable, and practical pavement defect detection process to support routine inspections, automatic alarms, and defect data retention.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] S1, Image Acquisition and Preprocessing: Acquire road inspection images, perform size unification, normalization and tensor quantization processing on the road inspection images, and generate standardized input feature tensors.

[0010] S2, Backbone Network Feature Extraction: The input feature tensor is input into the backbone network for multi-level feature extraction, outputting multi-scale feature maps. The backbone network uses the EfficientNetV2-S network, and in the shallow feature extraction stage of this backbone network, the Fused-MBConv module replaces the original MBConv module to enhance the extraction efficiency of shallow texture features and reduce memory access overhead.

[0011] S3, Neck Network Feature Fusion: The multi-scale feature maps are input into the SlimNeck network for cross-scale feature fusion to generate fused multi-scale features. The SlimNeck network's feature pyramid downsampling path includes a GSConv module to reduce computation while maintaining feature expressive power; and a VoVGSCSP module is introduced after multi-scale feature concatenation to enhance the collaborative expression of deep semantics and shallow spatial information.

[0012] S4, Detection Head Prediction and Loss Optimization: The fused multi-scale features are input into the detection head network, outputting candidate predicted bounding boxes and class confidence scores for road surface defects. During model training, the WIoUv3 loss function is used as the bounding box regression loss. The WIoUv3 loss function introduces a dynamic non-monotonic focusing mechanism based on outlier degree to dynamically adjust the contribution of different samples to gradient updates.

[0013] S5, Post-processing and Result Output: Post-processing is performed on the candidate predicted bounding boxes and category confidence scores. The post-processing includes using a non-maximum suppression algorithm to filter redundant candidate boxes, and finally outputting the classification results, confidence scores, and absolute coordinates of the bounding boxes for pavement defects.

[0014] The present invention also provides an improved YOLO11 road surface defect lightweight detection system for routine inspections, comprising: a data acquisition and preprocessing module, a lightweight backbone network module, an efficient feature fusion neck module, a detection head module, a loss calculation module, and a post-processing output module. Each module cooperates with the other to perform the steps of the above method.

[0015] In addition, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] By introducing the lightweight EfficientNetV2-S backbone network and the efficient SlimNeck feature fusion architecture, the number of parameters in this invention is reduced by 27.1%, and the floating-point operation volume (GFLOPs) is reduced by 38.4%, making it well-suited for deployment on embedded or mobile devices with limited computing power. Experimental results show that the inference speed of this invention reaches 208 FPS in an RTX 4090 experimental environment, which can fully meet the requirements of high frame rate and low latency for high-speed mobile inspection. While significantly reducing weight, the detection accuracy of the model has not decreased thanks to the optimization of the WIoUv3 loss function: the mAP50 on the Jilin-RDD dataset reaches 0.702, which is better than YOLOv5s and YOLOv8s, and the localization of difficult samples such as cracks is more accurate, with a significantly reduced false negative rate. Compared to large models such as RT-DETR, this invention avoids high computational redundancy; compared to general lightweight improvement schemes based on MobileNet, this invention retains stronger feature representation capabilities, thus achieving a good balance between accuracy, speed and computational power consumption, providing a preferred technical solution for routine road inspection. Attached Figure Description

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0019] Figure 1 This is a schematic diagram of the overall process of the lightweight detection method for pavement defects of the present invention.

[0020] Figure 2 This is a schematic diagram of the overall architecture of the improved YOLO11 model of the present invention, showing the connection relationship between the input end, the backbone network, the neck network and the detection head.

[0021] Figure 3 This is a schematic diagram of the detection results of the improved model of the present invention, showing the visualization effect of detection of different types of pavement defects.

[0022] Figure 4 This is a schematic diagram of the Fused-MBConv module structure.

[0023] Figure 5 This is a schematic diagram of the MBConv module structure, used for communication with... Figure 4 The improvements of this invention are explained in comparison.

[0024] Figure 6 This is a schematic diagram of the GSConv module structure, showing the combination of standard convolution, depthwise separable convolution, and channel shuffling.

[0025] Figure 7 This is a schematic diagram of the GSBottleneck module structure.

[0026] Figure 8 This is a schematic diagram of the VoVGSCSP module structure, showing the feature fusion path of cross-stage partial connections and primary aggregation.

[0027] Figure 9 This is a schematic diagram of the SPPF module structure, illustrating the role of spatial pyramid pooling in high-level feature enhancement. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0029] Unless otherwise stated, the relevant terms in this embodiment are defined as follows:

[0030] YOLO11: A single-stage object detection algorithm that refactors the object detection task into an end-to-end regression problem.

[0031] EfficientNetV2-S: A lightweight convolutional neural network optimized for mobile devices, employing the Fused-MBConv module to reduce memory access overhead and improve inference speed.

[0032] Fused-MBConv: A lightweight convolutional unit that combines extended convolution and depthwise separable convolution in traditional MBConv into a single standard convolutional operation, used to reduce memory access overhead and improve the efficiency of shallow texture feature extraction.

[0033] MBConv: A commonly used inverse residual convolution module in mobile networks, including expanded convolution, depthwise separable convolution, and projective convolution.

[0034] SlimNeck: A lightweight feature fusion structure for multi-scale object detection tasks, designed to improve the fusion efficiency between high-level semantic features and low-level spatial details.

[0035] GSConv: A lightweight convolution module that combines standard convolution, depthwise separable convolution, and channel shuffling operations to retain key feature information while reducing computational cost.

[0036] GSBottleneck: A bottleneck structure composed of GSConv as the basic unit, containing residual connections, used to achieve efficient feature transfer.

[0037] VoVGSCSP: A lightweight feature fusion module that combines the VoV aggregation concept with the CSP cross-stage partial connection concept, fusing deep semantic and shallow spatial information through a one-time aggregation strategy.

[0038] SPPF: Fast Spatial Pyramid Pooling Module, which expands the receptive field of the network and enhances the expression of contextual information through multi-scale pooling operations.

[0039] WIoUv3 (Wise-IoU v3): A bounding box regression loss function based on a dynamic non-monotonic focusing mechanism, which intelligently adjusts the contribution of different samples to gradient updates by constructing dynamic gradient gain coefficients.

[0040] Example 1: Specific Implementation of a Lightweight Detection Method for Road Surface Defects

[0041] This embodiment provides an improved lightweight detection method for YOLO11 pavement defects for routine inspections, such as... Figure 1 As shown, it includes the following steps:

[0042] Step 1: Frame extraction and basic image preprocessing of multi-source road surface video:

[0043] Acquire routine road inspection video streams collected by vehicle-mounted equipment. For inspection videos of different road sections, categorize and organize them according to road section information, and extract road image frames from the corresponding videos according to preset frame extraction rules;

[0044] Let the original inspection video sequence be:

[0045]

[0046] in, Indicates the first Frame road surface image, This indicates the total number of video frames.

[0047] Assume the original video capture frame rate is The target frame rate is Then the frame extraction interval can be expressed as:

[0048]

[0049] Based on the frame extraction interval, the image set after frame extraction It can be represented as:

[0050]

[0051] Using the above method, static road surface image samples for subsequent detection can be extracted from continuous inspection videos. Then, the images obtained by frame extraction are uniformly scaled to a preset standard size and normalized to convert them into a tensor format acceptable to deep learning models, generating preprocessed input feature tensors.

[0052] Let the standard input size be... The preprocessed input feature tensor can then be expressed as:

[0053]

[0054] The images obtained by frame extraction are uniformly scaled to 640×640 pixels and normalized (e.g., the pixel values ​​are divided by 255) and then converted into tensor format to generate the input feature tensor. .

[0055] After the above basic preprocessing, a standardized road surface image feature tensor that can be directly input into the improved detection model is obtained. The purpose of this step is to unify the input scale of image samples from different road sections, at different times, and under different acquisition conditions, so as to ensure that the input format of the subsequent backbone network and detection head network is consistent.

[0056] Step 2: Feature extraction based on the reconstructed backbone network

[0057] The input feature sheets are fed into the backbone network. In this embodiment, EfficientNetV2-S is used as the backbone network, and the original MBConv module is replaced with the Fused-MBConv module in its shallow feature extraction stage. The Fused-MBConv module merges the extended convolution and depthwise separable convolution in the traditional MBConv module into a single standard convolution operation, thereby reducing unnecessary intermediate operation overhead and improving the efficiency of shallow feature extraction while ensuring the expressive power of shallow features. Since many targets in road surface defect detection tasks exhibit thin cracks, irregular edges, or local texture abrupt changes, the preservation of shallow high-frequency texture information plays an important role in the subsequent detection accuracy.

[0058] Let the input feature tensor be The channel expansion ratio is Then the number of intermediate channels satisfies:

[0059]

[0060] The calculation process of the reconstructed Fused-MBConv is expressed as follows:

[0061]

[0062] in, This is a standard convolution operation used to extract high-frequency texture information in the shallow layer of the road surface, such as the edges of microcracks, fine-grained fracture areas, and abrupt changes in local texture information of defects. This is a projective convolution used to complete channel mapping and output feature integration.

[0063] By configuring the expansion ratio parameter, the generation of minor defect features can be accelerated while reducing memory access overhead. Preferably, the expansion ratio parameter is t=2 or t=4.

[0064] A smaller expansion ratio is more conducive to controlling model complexity, while a larger expansion ratio is more conducive to enhancing the representation ability of shallow features. The choice between the two can be made based on the requirements for lightweight model and detection accuracy.

[0065] After extraction by the backbone network, a multi-scale feature map is output, denoted as:

[0066]

[0067] in, , and These correspond to shallow texture features, mid-level structural features, and deep semantic features at different levels. Shallow features are useful for characterizing crack edges and fine-grained disease textures, mid-level features are useful for characterizing the local structural morphology of diseases, and deep features are useful for characterizing the overall semantic information of larger areas of diseases such as pits, network cracks, and repair cracks.

[0068] Step 3: Multi-scale feature pyramid fusion based on the SlimNeck architecture

[0069] The multi-scale feature maps output by the backbone network are input into the reconstructed SlimNeck network for cross-scale information aggregation, so as to realize information fusion between features at different levels and enhance the model's ability to jointly express pavement defects at different scales.

[0070] Because road surface defects exhibit significant scale differences in real-world scenarios—for example, microcracks typically appear as thin, elongated, and localized texture variations, while potholes, network cracks, and repair cracks usually manifest as large-scale structural anomalies—it is necessary to combine shallow spatial detail information with deep semantic information through multi-scale feature fusion to improve the detection capability of multi-scale defect targets.

[0071] In the feature pyramid downsampling path of the neck network, a GSConv module is configured to replace the standard downsampling convolution. Let the input high-dimensional features be... The expected number of output channels is The specific operation process of the GSConv module is as follows: First, local dense features are generated through standard convolution. :

[0072]

[0073] Subsequently, globally sparse redundant features are generated through depthwise separable convolution. :

[0074]

[0075] Finally, the locally dense features and the globally sparse and redundant features are concatenated along the channel dimension, and a channel shuffling operation is performed for recombination.

[0076]

[0077] in, This represents standard convolution, primarily used to preserve locally dense and effective feature responses; This represents depthwise separable convolution, used to generate sparse feature representations with low computational cost; This indicates a channel shuffling operation, used to facilitate the recombination and interaction of information between different channel features.

[0078] The GSConv module described above can balance the expressive power of convolutional features and lightweight characteristics to a certain extent, so that when the neck network performs downsampling feature transfer, it can maintain local structural information and reduce redundant computation caused by standard convolution.

[0079] After multi-scale feature concatenation, the VoVGSCSP module is introduced to fuse deep semantic and shallow spatial information through a one-time aggregation strategy. This module mainly improves the representation quality of the fused multi-scale features by uniformly aggregating and compressing the concatenated features, while maintaining feature reusability. Its computation process can be summarized as follows:

[0080]

[0081] in, This represents the intermediate features extracted after the lightweight bottleneck structure. This represents cross-stage direct connection branch features. The one-time aggregation method described above enhances the synergistic representation effect between deep and shallow features.

[0082] This step enhances the model's ability to perceive the spatial boundaries of pavement defects with varying scales by shuffling and unifying the channels of dense and sparse features, thereby obtaining a fused multi-scale feature output:

[0083]

[0084] in, , and These represent the three scales of fused features fed into the detection head network. Through the above processing, the subsequent detection head can jointly predict defects such as small cracks, complex cracks, and pits based on the fused features at different scales.

[0085] Step 4: Bounding box regression and dynamic parameter assignment based on outlier assessment

[0086] The fused multi-scale features are input into the detection head network, which outputs candidate predicted bounding boxes and class confidence scores for diseases. During model training, the WIoUv3 loss function is configured to calculate the regression loss between the predicted bounding boxes and the ground truth boxes. Compared to conventional IoU-type loss functions, WIoUv3 introduces a dynamic non-monotonic focusing mechanism based on outlier, enabling the model to dynamically adjust the contribution of different samples to gradient updates during training, thereby improving the ability to locate complex disease targets.

[0087] Define the bounding box regression loss as:

[0088]

[0089] in, Intersection over Union (IoU) loss is used to measure the degree of overlap between the predicted bounding box and the ground truth bounding box. This is a center point distance penalty term used to constrain the offset between the center position of the predicted bounding box and the center position of the ground truth bounding box; These are dynamic gradient gain coefficients constructed based on outlier degree, used to dynamically adjust the gradient weights of the current sample.

[0090] outlier and gain coefficient The calculation formula is:

[0091]

[0092]

[0093] in, The gradient separation IoU loss for the current anchor frame. This is the moving average of the losses across all anchor frames; and To adjust the hyperparameters of the focusing curve.

[0094] In addition to the bounding box regression loss, the detection head also outputs the category prediction result and the target confidence. The overall training loss can be expressed as:

[0095]

[0096] in, , and These represent the weight coefficients of the bounding box regression loss, classification loss, and confidence loss, respectively. Through this loss design, the model can achieve joint optimization in classification, localization, and target existence determination.

[0097] Step 5: Output and Post-processing of Detection Results

[0098] During the model inference phase, the detection head outputs a set of candidate bounding boxes and corresponding category confidence scores for the current road surface video frame.

[0099] Let the current frame have a total of If there are n candidate targets, then the initial candidate set can be represented as:

[0100]

[0101] in, Indicates the first 1 candidate bounding box Indicates candidate category labels, This represents the confidence score for the corresponding category.

[0102] To eliminate duplicate detection results and retain the optimal target box, a non-maximum suppression algorithm is used to filter redundant candidate boxes with an overlap exceeding a preset threshold. For any two candidate boxes... and Its intersection-union ratio is defined as:

[0103]

[0104] If the following conditions are met:

[0105]

[0106] and:

[0107]

[0108] Then delete candidate boxes with lower confidence. (in (This refers to the set overlap threshold). Through the above post-processing, the situation where the same disease target is repeatedly predicted can be effectively reduced, thereby improving the stability and accuracy of the final detection results.

[0109] After nonmaximum suppression, the final output set can be represented as:

[0110]

[0111] in, To determine the final number of targets to retain, each detection result includes: category Confidence level and the absolute coordinates of the bounding box .

[0112] The final output includes the target classification results, confidence scores, and precise absolute coordinates of the bounding boxes for five types of pavement defects. These five types of pavement defects include: transverse cracks, longitudinal cracks, mesh cracks, repair cracks, and potholes, thus completing automated target detection.

[0113] Example 2: Experimental Verification

[0114] To verify the effectiveness of the present invention, a comparative experiment was conducted on the Jilin-RDD hybrid dataset.

[0115] Dataset Construction: Data from RDD2024-China and Road-Comp competitions, along with on-site inspection images of the Jilin Expressway, were integrated to construct a dataset containing 3,140 samples, covering five types of defects: transverse cracks, longitudinal cracks, mesh cracks, repair cracks, and potholes. The training set to test set ratio was 8:2.

[0116] Experimental environment and configuration: RTX 4090 GPU, PyTorch 2.5.1 framework. Input image size 640×640, Batch Size set to 16, training for 300 epochs. SGD optimizer was used, with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005.

[0117] Comparative Experimental Results: Parameter Quantity and Computational Cost: The model of this invention has 6.87M parameters and 13.3 GFLOPs, both the lowest among the comparative models. Compared with YOLOv8s, the parameter quantity is reduced by 27.1% and GFLOPs are reduced by 53.2%.

[0118] Detection accuracy: The mAP50 of the model of this invention is 0.702, which is 3.9% higher than YOLOv5s, 1.4% higher than YOLOv8s, and basically the same as or even slightly better than the baseline YOLO11s (0.701).

[0119] Inference speed: The model of this invention achieved an inference speed of 208 FPS in the RTX 4090 experimental environment, which can smoothly process video stream data.

[0120] Visualization of detection results: such as Figure 3 As shown, the present invention can accurately locate minute longitudinal / transverse cracks and large-area mesh cracks, and the generated bounding box has a high degree of fit, with no obvious missed detections or false detections, proving its robustness in complex scenarios.

[0121] Module structure description: combined with Figures 4 to 9 The key modules used in this invention will be described in detail below:

[0122] Fused-MBConv module ( Figure 4 : The extended convolution (1×1 convolution) and the depthwise separable convolution (3×3 depthwise convolution) are merged into a standard convolution (3×3 convolution), and then passed through BatchNorm, SiLU activation, projective convolution (1×1 convolution), BatchNorm and Dropout in sequence, and finally residual connection with the input (when the stride is 1 and the number of channels matches).

[0123] MBConv module ( Figure 5 The traditional structure consists of a 1×1 extended convolution, a 3×3 depthwise separable convolution, a SE attention module, and a 1×1 projective convolution. In comparison, the Fused-MBConv module reduces memory accesses and is more suitable for shallow texture extraction.

[0124] GSConv module ( Figure 6The input features are first processed by standard convolution to generate dense features for some channels, and then processed by depthwise separable convolution to generate sparse features for the remaining channels. After concatenation, channel shuffling is performed to achieve lightweight and efficient feature transformation.

[0125] GSBottleneck module ( Figure 7 ): It consists of two GSConv modules, which may contain residual connections in between, forming a bottleneck structure, used to deepen the network without significantly increasing the computational cost.

[0126] VoVGSCSP module ( Figure 8 The input features are divided into two paths. One path passes through multiple GSBottlenecks to extract deep features, while the other path is directly connected as cross-stage features. Finally, the two features are concatenated and fused through a convolution to output the final feature.

[0127] SPPF module ( Figure 9 The input features are subjected to three 5×5 max pooling operations in sequence. The original features and the results of the three pooling operations are concatenated along the channel dimension and then output through a convolutional layer to expand the receptive field.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various improvements and modifications without departing from the spirit and principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An improved lightweight detection method for YOLO11 pavement defects for routine inspections, comprising the following steps: S1: Acquire road inspection images and preprocess the road inspection images to generate standardized input feature tensors; S2: Input the input feature tensor into the backbone network for multi-level feature extraction and output multi-scale feature maps; the backbone network is an EfficientNetV2-S network, and the shallow feature extraction stage of the backbone network uses the Fused-MBConv module instead of the original MBConv module of the network. S3: Input the multi-scale feature map into the reconstructed SlimNeck network for cross-scale feature fusion to generate fused multi-scale features; wherein, the feature pyramid downsampling path of the SlimNeck network is configured with a GSConv module, and a VoVGSCSP module is introduced after the multi-scale features are concatenated. S4: Input the fused multi-scale features into the detection head network and output the candidate predicted bounding boxes and category confidence of road surface defects; during the model training stage, the WIoUv3 loss function is used as the bounding box regression loss, and the WIoUv3 loss function introduces a dynamic non-monotonic focusing mechanism based on outlier. S5: Perform post-processing on the candidate predicted bounding boxes and category confidence scores, and output the final pavement distress detection results; the post-processing includes filtering redundant candidate boxes using a non-maximum suppression algorithm.

2. The improved YOLO11 pavement distress lightweight detection method for routine inspection as described in claim 1, characterized in that, The process described in S2, which involves inputting the input feature tensor into the backbone network for multi-level feature extraction and outputting multi-scale feature maps, includes: Let the input feature tensor be The channel expansion ratio is Then the number of intermediate channels satisfies: ； The calculation process of the Fused-MBConv module is represented as follows: ； in, This is a standard convolution operation used to extract shallow high-frequency texture information from road surfaces. For projective convolution; Configure the expansion ratio parameter to reduce memory access overhead and accelerate the generation of minor disease features; the expansion ratio is selected based on the model's lightweight requirements and detection accuracy requirements. After extraction by the backbone network, a multi-scale feature map is output, denoted as: ； in, , and These correspond to shallow texture features, mid-level structural features, and deep semantic features at different levels.

3. The improved YOLO11 road surface defect lightweight detection method for routine inspection as described in claim 1, characterized in that, The calculation process of the GSConv module described in S3 is as follows: Let the input high-dimensional features be The expected number of output channels is The specific operation process of the GSConv module is as follows: generating local dense features through standard convolution. : ； Global sparse redundant features are generated through depthwise separable convolution. : ； The locally dense features and the globally sparse and redundant features are concatenated along the channel dimension, and then a channel shuffling operation is performed for recombination: ； in, Represents standard convolution. This represents depthwise separable convolution; This indicates a channel rinsing operation.

4. The improved YOLO11 pavement distress lightweight detection method for routine inspection as described in claim 1, characterized in that, The VoVGSCSP module described in S3 is configured to fuse deep semantic and shallow spatial information through a one-time aggregation strategy. The computation process is as follows: ； in, This represents the intermediate features extracted after the lightweight bottleneck structure. This indicates a cross-stage direct-connection branch feature; Output the fused multi-scale features: ； in, , and These represent the three scale fusion features fed into the detection head network.

5. The improved YOLO11 pavement distress lightweight detection method for routine inspection as described in claim 1, characterized in that, The WIoUv3 loss function in S4 is defined as follows: ； in, , and These represent the weight coefficients of the bounding box regression loss, classification loss, and confidence loss, respectively. Among them, the bounding box regression loss Defined as: ； in, To compare the losses; The penalty term is the distance from the center point; These are the dynamic gradient gain coefficients constructed based on outlier degree; outlier and gain coefficient The calculation formula is: ； ； in, The gradient separation IoU loss for the current anchor frame. This is the moving average of the losses across all anchor frames; and To adjust the hyperparameters of the focusing curve.

6. The improved YOLO11 pavement distress lightweight detection method for routine inspection as described in claim 1, characterized in that, The overlap judgment condition of the nonmaximum suppression algorithm in S5 is: Let the current frame have a total of If there are n candidate targets, then the initial candidate set can be represented as: ； in, Indicates the first Candidate bounding boxes, Indicates candidate category labels, This represents the confidence score for the corresponding category; For any two candidate boxes and Its intersection-union ratio is defined as: ； If the following conditions are met: ； and: ； Then delete candidate boxes with lower confidence. ,in The set overlap threshold.

7. The improved YOLO11 pavement distress lightweight detection method for routine inspection as described in claim 1, characterized in that, The pavement distress detection output in S5 includes the pavement distress classification results, confidence level, and absolute coordinates of the bounding box; The road surface defects include: transverse cracks, longitudinal cracks, network cracks, repair cracks, and potholes.

8. An improved YOLO11 lightweight pavement defect detection system for routine inspections, used to implement the method described in any one of claims 1 to 8, characterized in that, include: Data acquisition and preprocessing module: used to acquire and preprocess road inspection images and generate input feature tensors; Lightweight backbone network module: built on EfficientNetV2-S, used to extract multi-scale feature maps from the input feature tensor; the shallow layers of the backbone network include Fused-MBConv modules; High-efficiency feature fusion neck module: built on the SlimNeck architecture, used to perform cross-scale fusion of the multi-scale feature maps; the downsampling path of the neck module includes a GSConv module, and a VoVGSCSP module is provided after feature concatenation; Detection head module: used to output candidate predicted bounding boxes and class confidence scores based on the fused multi-scale features; Loss calculation module: used to calculate bounding box regression loss using the WIoUv3 loss function during the training phase; Post-processing output module: used to filter candidate predicted bounding boxes through non-maximum suppression and output the final detection results.

9. A 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 steps of the method for detecting lightweight road surface defects as described in any one of claims 1 to 8.

Images