YOLOv11 road disease detection method and system based on multi-scale feature enhancement

Abstract

The invention discloses a YOLOv11 road disease detection method and system based on multi-scale feature enhancement. The method comprises the steps of road image acquisition and marking, adaptive learning rate adjustment and data enhancement, construction of a YOLOv11 model embedded with a sub-pixel level edge enhancement module and a wavelet transform feature extraction module, staged training, model evaluation, disease detection and the like. Wherein the sub-pixel-level edge enhancement module is used for enhancing fine crack features through multi-operator fusion and a sub-pixel convolution technology; the wavelet transformation module improves the multi-scale feature perception capability through adaptive wavelet basis selection and a multi-stage decomposition and reconstruction mechanism. The system correspondingly comprises a data preprocessing module, a model building module, a training optimization module and an evaluation deployment module. On the basis of keeping the real-time performance of the YOLOv11, the detection precision and robustness of multi-scale diseases in low-contrast, sub-pixel-level cracks and complex environments are remarkably improved, and the method is suitable for road maintenance and safety monitoring scenes.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent transportation infrastructure monitoring technology, specifically to a YOLOv11 road defect detection method and system based on multi-scale feature enhancement. Background Technology

[0002] With the continuous expansion of my country's highway network and the increase in its service life, road surfaces are gradually developing various defects under the combined effects of vehicle loads, environmental factors, and material aging. These defects not only affect driving comfort but also pose a potential threat to road traffic safety. Traditional road inspections mainly rely on manual patrols, which have drawbacks such as low efficiency, strong subjectivity, and significant safety hazards, making it difficult to meet the needs of modern road maintenance.

[0003] In recent years, computer vision-based automatic detection technology has made significant progress. Especially in the field of deep learning, the YOLO series of algorithms has been widely used in road defect detection due to its excellent real-time performance. YOLOv11, as the latest version of this series, has improved in both detection speed and accuracy. However, in real-world road scenarios, this algorithm still faces two key technical challenges: First, linear cracks and other road defects typically exhibit low contrast and sub-pixel width, making it easy for traditional convolutional neural networks to lose this subtle edge information during feature extraction, resulting in a high false negative rate. Second, the complexity of the road environment (such as changes in lighting, shading, and road surface texture interference) poses significant challenges to the accurate extraction of defect features. Existing algorithms lack the ability to perceive multi-scale features and struggle to effectively handle defect targets of varying sizes.

[0004] Although some studies have attempted to introduce multi-scale analysis methods such as wavelet transform into target detection, most of them only remain at the level of simple feature replacement, failing to fully leverage the advantages of wavelet analysis in multi-scale feature extraction and detail preservation, and lacking specific optimization for road damage characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a YOLOv11 road defect detection method and system based on multi-scale feature enhancement, so as to solve the problems of low efficiency, strong subjectivity and great safety hazards in traditional road detection, which mainly relies on manual inspection.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a YOLOv11 road defect detection method based on multi-scale feature enhancement, comprising the following steps: Step 1: Acquire road surface images and store them in the road disease dataset. Label the diseased areas in the road surface images in the road disease dataset to obtain image data. Step 2: Through adaptive learning rate adjustment and data augmentation techniques, the image data in the road disease dataset is optimized to obtain preprocessed image data; Step 3: Construct a YOLOv11 detection model, embedding a sub-pixel-level edge enhancement module and a wavelet transform feature extraction module into the YOLOv11 framework; Step 4: Adopt a phased training strategy. First, use a large natural image dataset in the public domain to pre-train the YOLOv11 detection model; then use the preprocessed image data obtained from the optimization in Step 2 to train the YOLOv11 detection model. Step 5: Evaluate the performance of the trained YOLOv11 detection model using the test set; Step 6: Use the trained YOLOv11 detection model to detect defects in the road surface image under test.

[0007] To optimize the above technical solution, the specific measures also include: In step 2, the adaptive learning rate adjustment refers to: Cosine annealing is used for warm-up and adaptive learning rate adjustment: the learning rate slowly decreases from the initial value to the minimum value according to the cosine function curve; at the end of each cosine cycle, the learning rate will suddenly restart and then start a new round of cosine decrease; combined with monitoring the validation set loss or accuracy, when the performance reaches a plateau, additional adjustment of the learning rate or periodic restart can be automatically triggered. Furthermore, in step 2, the data augmentation techniques include: geometric transformation, color and lighting perturbation, blending enhancement, and noise and occlusion simulation.

[0008] In step 3, a subpixel-level edge enhancement module is embedded in the YOLOv11 framework, as follows: Using a multi-branch parallel structure, the preprocessed image data is initially processed through Backbone convolution to obtain an intermediate feature map. Multiple different edge detection operators are used to extract edge features from the input intermediate feature map, resulting in multiple edge feature maps. Through an adaptive weight fusion mechanism, the multiple edge feature maps are weighted and fused to obtain a fused edge feature map. Subpixel convolution upsampling is performed on the fused edge feature map to generate an edge response map. Based on the edge response map, an edge attention map is generated through Gaussian smoothing and the Sigmoid function. The edge attention map is then multiplied by the original input feature map to output an enhanced feature map.

[0009] Furthermore, the adaptive weight fusion mechanism is implemented through the following steps: The edge feature maps extracted from each branch are concatenated along the channel dimension; global average pooling is performed on the concatenated feature maps to obtain a channel descriptor vector; the channel descriptor vector is input into a lightweight network consisting of fully connected layers to generate initial weight scores for each edge operator; the initial weight scores are normalized using Softmax to obtain adaptive fusion weights with a weight sum of 1.

[0010] Furthermore, in step 3, a wavelet transform feature extraction module is embedded in the YOLOv11 framework, as follows: An adaptive selection mechanism is used to dynamically select the optimal wavelet basis function for the input feature map. The selected wavelet basis is used to perform multi-level wavelet decomposition on the input feature map to obtain sub-bands containing low-frequency information and some high-frequency information. Through a learnable fusion weight mechanism, the features of each sub-band after decomposition are adaptively weighted and fused. The fused features are reconstructed by wavelet and added to the original input features through residual connections and depthwise separable convolution to output the final multi-scale features.

[0011] Step 5, specifically evaluating the performance of the trained YOLOv11 detection model on an independent test set, includes: On an independent test set, the precision, recall, and average accuracy of the YOLOv11 detection model for various types of road defects, as well as the crossover ratio threshold, were calculated.

[0012] As another important technical solution, the present invention also provides a YOLOv11 road defect detection system based on multi-scale feature enhancement, comprising: The data acquisition and preprocessing module is used to acquire road surface images and store them in the road disease dataset. It annotates the diseased areas of the road surface images in the road disease dataset to obtain image data. Through adaptive learning rate adjustment and data augmentation technology, it optimizes the image data in the road disease dataset to obtain preprocessed image data. The model building module is used to build a YOLOv11 detection model, in which a sub-pixel-level edge enhancement module and a wavelet transform feature extraction module are embedded in the YOLOv11 framework. The model training and optimization module is used to adopt a phased training strategy. First, the YOLOv11 detection model is pre-trained using a large natural image dataset in the public domain. Then, the YOLOv11 detection model is trained using the preprocessed image data obtained from step 2. The model evaluation and deployment module is used to evaluate the performance of the trained YOLOv11 detection model on the test set; and to perform defect detection on the road surface image under test using the trained YOLOv11 detection model.

[0013] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the YOLOv11 road defect detection method based on multi-scale feature enhancement as described above.

[0014] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute a YOLOv11 road defect detection method based on multi-scale feature enhancement as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. The SEEM module proposed in this invention is specifically designed for the sub-pixel characteristics of road cracks. Through multi-operator fusion and sub-pixel reconstruction technology, it significantly improves the detection capability of low-contrast linear cracks and solves the problem of missed detection by traditional methods on such targets.

[0016] 2. The wavelet transform feature extraction module of this invention goes beyond simple feature replacement. Through adaptive wavelet basis selection, multi-level decomposition and reconstruction, and learnable fusion mechanism, it fully leverages the advantages of wavelet analysis in multi-scale feature extraction, providing more effective feature representation for road defect detection.

[0017] 3. The dual-module collaborative working mechanism of the present invention enables the model to simultaneously possess the ability to preserve details and the ability to perceive multiple scales. It can capture subtle crack features and handle disease targets of different sizes, exhibiting stronger robustness in complex road environments.

[0018] 4. While maintaining the original real-time performance of YOLOv11, this invention significantly improves the detection accuracy, especially in detecting minute cracks and in complex environments, providing more reliable technical support for practical road maintenance applications.

[0019] 5. The module design of this invention is versatile and can be applied not only to road defect detection but also to other visual detection tasks that require fine feature extraction, thus having broad application prospects. Attached Figure Description

[0020] Figure 1 : Overall flowchart of the present invention. Figure 2 Diagram of the improved YOLOv11 network structure of this invention. Figure 3 : Structural diagram of the subpixel-level edge enhancement module of the present invention. Figure 4 The flowchart of the innovative wavelet transform feature extraction module of this invention. Figure 5: Schematic diagram of the multi-level wavelet decomposition and reconstruction strategy of this invention. Figure 6 : Schematic diagram of the adaptive wavelet basis selection mechanism of this invention. Detailed Implementation The present invention will be further described in detail below through specific embodiments, but it should not be construed as limiting the scope of the subject matter of the present invention to the following embodiments. All technologies implemented based on the above content of the present invention fall within the scope of the present invention.

[0021] This invention provides a YOLOv11 road defect detection method based on multi-scale feature enhancement, comprising the following steps: Step 1: Acquire road surface images and store them in the road disease dataset. Label the diseased areas in the road surface images in the road disease dataset to obtain image data. In some implementations, road surface images are acquired using vehicle-mounted acquisition equipment to construct a road defect dataset containing various typical defects. Professional annotation tools are then used to finely annotate the defect areas in the images, with annotation categories including at least the main defect types such as longitudinal cracks, transverse cracks, network cracks, and potholes.

[0022] Step 2: Through adaptive learning rate adjustment and data augmentation techniques, the image data in the road disease dataset is optimized to obtain preprocessed image data; In some implementations, to dynamically optimize the training process, we employ a cosine annealing warm-up strategy as the core of adaptive learning rate adjustment: Cosine annealing: The learning rate decreases slowly from its initial value to its minimum value according to a cosine function curve. This smooth descent helps the model escape sharp local optima towards the end of training and converge to a flatter minimum, thereby improving generalization ability.

[0023] Warm-up Reset: At the end of each cosine cycle, the learning rate is suddenly "reset" to a higher value (not starting from zero), and then a new round of cosine descent begins. This "reset" mechanism helps the model escape the current local optima and continue to explore a better solution space.

[0024] Adaptive adjustment: By combining monitoring of validation set loss or accuracy, when performance reaches a plateau, additional adjustments to the learning rate or periodic restarts can be automatically triggered, achieving true "adaptiveness".

[0025] Preferably, to simulate complex and ever-changing real-world road scenarios and enhance model robustness, we introduced an enhanced data augmentation pipeline during training: Geometric transformations include random horizontal / vertical flips, random rotations (-10° to +10°), random scaling (0.8 to 1.2x), and random pruning to improve the model's invariance to disease orientation and scale.

[0026] Color and lighting perturbation: Adjust the brightness, contrast, saturation, and hue of the image to simulate different weather and lighting conditions (such as overexposure, shadows, and dusk) to ensure that the model does not depend on specific color information.

[0027] Advanced hybrid enhancements include Mosaic enhancements and MixUp enhancements: Mosaic enhancement: Four training images are stitched together into one for training, which greatly enriches the background information and target context of a single image, and improves the detection capability of small target diseases.

[0028] MixUp enhancement: This technique linearly blends two images with certain weights, and their labels are also blended accordingly. This regularization technique encourages the model to learn smoother decision boundaries, reducing overfitting.

[0029] Noise and Occlusion Simulation: Randomly add Gaussian noise, motion blur, or simulate partial occlusion to improve the model's stability under low-quality images or partial occlusion conditions.

[0030] Step 3: Construct a YOLOv11 detection model, embedding a sub-pixel-level edge enhancement module and a wavelet transform feature extraction module into the YOLOv11 framework; In road defect detection tasks, targets such as linear cracks often exhibit low contrast and sub-pixel width characteristics. Traditional convolutional neural networks are prone to losing these subtle edge information during feature extraction. Although the C3k2 module enhances gradient flow through cross-stage partial connections, its kernel size is limited and the feature mixing depth between channels is insufficient. It still struggles to process such subtle features, leading to the loss of edge and texture details in complex backgrounds or scenes with dense multiple targets.

[0031] The core innovation of the sub-pixel-level edge enhancement module SEEM in this invention lies in its multi-branch parallel architecture, integrating multiple edge detection operators (Sobel, Laplacian, Prewitt, etc.) to form a complementary edge feature extraction mechanism. This significantly improves the response intensity to sub-pixel-level crack edges and effectively suppresses background noise interference. Furthermore, it introduces sub-pixel convolution (Pixel-Shuffle) technology to upsample and reconstruct the fused edge features. Specifically, it first integrates the edge features extracted from multiple branches using a weighted fusion strategy, where the weight parameters can be learned through the network.

[0032] In some implementations, efficient multi-angle edge feature extraction is achieved through 1×1 convolution, which not only ensures computational efficiency but also fully extracts multi-angle edge features, laying a high-quality foundation for subsequent feature fusion.

[0033] The preprocessed image data is initially processed using Backbone convolution to obtain intermediate feature maps, which are then input into the intermediate feature maps. X The process proceeds in parallel through n different edge detection branches. Each branch simulates the operation of the operator through a 1×1 convolution, generating a corresponding edge feature map. , where i = 1, 2, ..., n.

[0034] The edge feature maps output by each branch are integrated through an adaptive weight fusion mechanism, and the fusion weights are dynamically learned through network training.

[0035] Edge feature maps of all branches output By concatenating the features along the channel dimension, a comprehensive feature map is obtained. .

[0036] right Global average pooling is performed to compress the spatial information of each channel into a global scalar value. This step transforms the feature map into a channel descriptor vector. This vector captures global statistics for each edge feature channel.

[0037] The vector z is input into a lightweight multilayer perceptron (consisting of two fully connected layers with a non-linear activation function such as ReLU in between). The network learns the non-linear relationships between channels and outputs an initial weight score vector.

[0038] Finally, the weight scores are normalized using the Softmax function to ensure that the sum of the weights of all branches is 1. This yields a set of adaptive fusion weights. ,(in Each weight corresponds to the importance of an edge operator branch.

[0039] Each edge feature map Fi Its corresponding adaptive weights wi Multiply them, then sum all the weighted feature maps to obtain the final fused comprehensive edge features. :

[0040] Subpixel convolution operations are used to upsample the fused edge features by a factor of 2 or 4 to generate a high-resolution edge response map, which is mathematically expressed as follows:

[0041] in, For the adaptive weights of each edge operator, This represents the i-th edge extraction operation.

[0042] This step solves the problem of insufficient feature representation caused by the small scale of subpixel-level cracks, providing key support for subsequent accurate detection.

[0043] To further enhance the robustness of edge features, Gaussian smoothing is applied to the high-resolution edge response map to eliminate noise interference and highlight the edge response of the disease, thereby significantly reducing the false negative rate of linear cracks. Subsequently, an edge attention map is generated using the Sigmoid activation function.

[0044] Finally, feature enhancement is achieved by multiplying the edge attention map with the original features through a feature modulation mechanism:

[0045] This step precisely enhances the feature signals in the crack edge region while suppressing irrelevant background information, significantly improving the distinction between the crack and the background.

[0046] The SEEM module can be flexibly embedded in key locations such as Stage 2 and Stage 3 of the YOLOv11 backbone network, or inserted after feature maps at various scales in the neck network, significantly improving the model's ability to perceive sub-pixel-level crack features. At the same time, due to its overall lightweight design, it maintains low computational overhead, perfectly adapting to the real-time requirements of road detection scenarios.

[0047] In some implementations, the wavelet transform feature extraction module of this invention introduces an adaptive wavelet basis selection mechanism. Unlike traditional methods that fixate on using Haar or Daubechies wavelet bases, this module dynamically analyzes the input feature map using a lightweight neural network and selects the optimal basis function from a preset wavelet basis library. Mathematically, this is expressed as:

[0048] in, To select a network, The wavelet basis matrix is ​​used. Network selection is performed. The output is normalized by the Softmax function to generate a probability distribution, which is used to weight and combine the basis functions in the wavelet basis library to obtain the optimal wavelet basis corresponding to the current input feature. .

[0049] This step significantly improves the model's adaptability to complex road environments and varied disease morphologies, solving the problem of insufficient feature extraction adaptability caused by traditional methods.

[0050] Preferably, a multi-level wavelet decomposition and reconstruction strategy is designed. The wavelet transform feature extraction module supports 2-3 levels of wavelet decomposition, with each level generating four sub-bands: LL, LH, HL, and HH. The low-frequency sub-band LL carries the main shape information and is used to expand the receptive field; the high-frequency sub-bands (LH, HL, HH) highlight detailed features through a selective enhancement mechanism. The reconstruction process uses inverse wavelet transform, but a learnable reconstruction filter is added to enhance flexibility, enabling the module to effectively capture the global morphological information of road defects (such as crack direction and pothole outline) and local detailed features (such as crack branching and texture changes).

[0051] The LL subband carries the main shape information to expand the receptive field and enhance the perception of the overall structure of the disease. The high-frequency subband is selectively enhanced and then adaptively fused with the LL subband to form a comprehensive feature representation with rich details and contextual information, which is particularly suitable for the identification of complex morphologies such as network cracks.

[0052] Learnable sub-band fusion weights are introduced into the wavelet transform feature extraction module, overcoming the limitations of traditional wavelet transform sub-band fusion that relies on manual experience. This enables adaptive optimization of the feature combination ratio, providing more accurate feature representations for lesions of different sizes and shapes. Each sub-band feature is first processed through a 1×1 convolution for channel adjustment, and then the fusion weights are calculated using an attention mechanism.

[0053] The final output feature is a weighted fusion of features from each sub-band:

[0054] The wavelet transform feature extraction module employs a residual connection mechanism, adding the original input to the wavelet transform output to ensure smooth gradient propagation while maintaining feature integrity.

[0055] in This represents a depthwise separable convolution operation. This operation is used to reconstruct the features before adding the residuals. Performing slight secondary feature fusion and channel adjustments can further improve performance without significantly increasing the number of parameters. Further enhance the expressive power of features.

[0056] This module primarily replaces the standard convolutional layers in Stage 1 and Stage 2 of the YOLOv11 backbone network, enabling multi-scale feature extraction in the shallow layers of the network. This effectively expands the receptive field and enhances the model's ability to perceive the global morphology and detailed features of cracks. Compared with existing technologies, this module not only achieves simple feature replacement but also, through adaptive mechanisms and learnable parameters, allows the module to dynamically adjust the extraction strategy based on specific disease characteristics. This improves the model's adaptability to complex road environments, fully leverages the advantages of wavelet analysis in multi-scale feature extraction, provides more effective feature representations for road disease detection, and significantly improves the detection accuracy for disease targets of different sizes.

[0057] Step 4: Adopt a phased training strategy. First, use a large natural image dataset in the public domain to pre-train the YOLOv11 detection model; then use the preprocessed image data obtained from the optimization in Step 2 to train the YOLOv11 detection model. In some implementations, pre-trained weights on large, general-purpose natural image datasets such as ImageNet or COCO are used to initialize the backbone network of the model.

[0058] The YOLOv11 detection model is trained using the preprocessed image data obtained from step 2. A small learning rate (usually 1 / 10 of the pre-training stage) is used for training, so that the model can adaptively learn the specific features of road defects without destroying the existing general feature extraction capabilities.

[0059] Step 5: Evaluate the performance of the trained YOLOv11 detection model using the test set; On an independent test set, the precision, recall, and average accuracy of the YOLOv11 detection model for various road defects, as well as the average accuracy with a set crossover ratio threshold, were evaluated.

[0060] Step 6: Use the trained YOLOv11 detection model to detect defects in the road surface image under test.

[0061] Example 1: Model Training and Testing Based on RDD Datasets Dataset preparation: High-quality road images were selected from the RDD2022 dataset and reclassified and labeled to include four main types of defects: longitudinal cracks, transverse cracks, network cracks, and potholes. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio.

[0062] Model configuration: Based on the YOLOv11 framework, the SEEM module is embedded in Stage 2 and Stage 3 of the Backbone network, and an innovative wavelet transform module is used to replace the traditional convolutional layers in Stage 1 and Stage 2. Training hyperparameter settings: 300 iterations, batch size 16, initial learning rate 0.01, and a cosine annealing learning rate scheduling strategy.

[0063] Training process: The model is initialized using pre-trained weights, and various data augmentation techniques such as random rotation, color jitter, and mosaic enhancement are employed. During training, the loss function and validation set performance are monitored, and a model snapshot is saved every 50 epochs.

[0064] Results analysis: Evaluation results on the test set show that the method of the present invention achieves an mAP@0.5 of 62.3% while maintaining a high detection speed (≥45FPS), which is 4.8 percentage points higher than the original YOLOv11, especially with a significant improvement in the detection accuracy of crack-type defects.

[0065] Example 2: Real-world road scenario test A field test was conducted on a section of a highway, using a detection vehicle equipped with a high-definition camera to collect images of 100 kilometers of road. The test results show that the method of this invention can effectively detect cracks wider than 0.2 mm, with a detection accuracy of 95.2% and a false detection rate controlled below 3.1%, meeting the requirements of practical engineering applications.

[0066] In another embodiment of the present invention, a YOLOv11 road defect detection system based on multi-scale feature enhancement is proposed, comprising: The data acquisition and preprocessing module is used to acquire road surface images and store them in the road disease dataset. It annotates the diseased areas of the road surface images in the road disease dataset to obtain image data. Through adaptive learning rate adjustment and data augmentation technology, it optimizes the image data in the road disease dataset to obtain preprocessed image data. The model building module is used to build a YOLOv11 detection model, in which a sub-pixel-level edge enhancement module and a wavelet transform feature extraction module are embedded in the YOLOv11 framework. The model training and optimization module is used to adopt a phased training strategy. First, the YOLOv11 detection model is pre-trained using a large natural image dataset in the public domain. Then, the YOLOv11 detection model is trained using the preprocessed image data obtained from step 2. The model evaluation and deployment module is used to evaluate the performance of the trained YOLOv11 detection model on the test set; and to perform defect detection on the road surface image under test using the trained YOLOv11 detection model.

[0067] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the YOLOv11 road defect detection method based on multi-scale feature enhancement as described above.

[0068] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute a YOLOv11 road defect detection method based on multi-scale feature enhancement as described above.

[0069] In another embodiment of the present invention, an electronic device is proposed, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements a YOLOv11 road defect detection method based on multi-scale feature enhancement as described above.

[0070] In another embodiment of the present invention, a computer-readable storage medium is provided storing a computer program that causes a computer to execute a YOLOv11 road defect detection method based on multi-scale feature enhancement as described above.

[0071] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent substitutions, and improvements made by those skilled in the art to the above embodiments without departing from the scope of the technical solution of the present invention, based on the technical essence of the present invention, shall still fall within the protection scope of the technical solution of the present invention.

Claims

1. A YOLOv11 road defect detection method based on multi-scale feature enhancement, characterized in that, Includes the following steps: Step 1: Acquire road surface images and store them in the road disease dataset. Label the diseased areas in the road surface images in the road disease dataset to obtain image data. Step 2: Through adaptive learning rate adjustment and data augmentation techniques, the image data in the road disease dataset is optimized to obtain preprocessed image data; Step 3: Construct a YOLOv11 detection model, embedding a sub-pixel-level edge enhancement module and a wavelet transform feature extraction module into the YOLOv11 framework; Step 4: Adopt a phased training strategy. First, use a large natural image dataset in the public domain to pre-train the YOLOv11 detection model; then use the preprocessed image data obtained from the optimization in Step 2 to train the YOLOv11 detection model. Step 5: Evaluate the performance of the trained YOLOv11 detection model using the test set; Step 6: Use the trained YOLOv11 detection model to detect defects in the road surface image under test.

2. The YOLOv11 road defect detection method based on multi-scale feature enhancement according to claim 1, characterized in that: In step 2, the adaptive learning rate adjustment refers to: Cosine annealing is used for warm-up and adaptive learning rate adjustment: the learning rate slowly decreases from the initial value to the minimum value according to the cosine function curve; at the end of each cosine cycle, the learning rate will suddenly restart and then start a new round of cosine decrease; combined with monitoring the validation set loss or accuracy, when the performance reaches a plateau, additional adjustment of the learning rate or periodic restart can be automatically triggered. In step 2, the data augmentation techniques include: geometric transformation, color and lighting perturbation, blending enhancement, and noise and occlusion simulation.

3. The YOLOv11 road defect detection method based on multi-scale feature enhancement according to claim 1, characterized in that: In step 3, a subpixel-level edge enhancement module is embedded in the YOLOv11 framework, as follows: Using a multi-branch parallel structure, the preprocessed image data is initially processed through Backbone convolution to obtain an intermediate feature map. Multiple different edge detection operators are used to extract edge features from the input intermediate feature map, resulting in multiple edge feature maps. Through an adaptive weight fusion mechanism, the multiple edge feature maps are weighted and fused to obtain a fused edge feature map. Subpixel convolution upsampling is performed on the fused edge feature map to generate an edge response map. Based on the edge response map, an edge attention map is generated through Gaussian smoothing and the Sigmoid function. The edge attention map is then multiplied by the original input feature map to output an enhanced feature map.

4. The YOLOv11 road defect detection method based on multi-scale feature enhancement according to claim 3, characterized in that: The adaptive weight fusion mechanism is implemented through the following steps: The edge feature maps extracted from each branch are concatenated along the channel dimension; global average pooling is performed on the concatenated feature maps to obtain a channel descriptor vector; the channel descriptor vector is input into a lightweight network consisting of fully connected layers to generate initial weight scores for each edge operator; the initial weight scores are normalized using Softmax to obtain adaptive fusion weights with a weight sum of 1.

5. The YOLOv11 road defect detection method based on multi-scale feature enhancement according to claim 1, characterized in that: In step 3, a wavelet transform feature extraction module is embedded into the YOLOv11 framework, as follows: An adaptive selection mechanism is used to dynamically select the optimal wavelet basis function for the input feature map. The selected wavelet basis is used to perform multi-level wavelet decomposition on the input feature map to obtain sub-bands containing low-frequency information and some high-frequency information. Through a learnable fusion weight mechanism, the features of each sub-band after decomposition are adaptively weighted and fused. The fused features are reconstructed by wavelet and added to the original input features through residual connections and depthwise separable convolution to output the final multi-scale features.

6. The YOLOv11 road defect detection method based on multi-scale feature enhancement according to claim 1, characterized in that: Step 5, specifically evaluating the performance of the trained YOLOv11 detection model on an independent test set, includes: On an independent test set, the precision, recall, and average accuracy of the YOLOv11 detection model for various types of road defects, as well as the crossover ratio threshold, were calculated.

7. A YOLOv11 road defect detection system based on multi-scale feature enhancement, characterized in that, include: The data acquisition and preprocessing module is used to acquire road surface images and store them in the road disease dataset. It annotates the diseased areas of the road surface images in the road disease dataset to obtain image data. Through adaptive learning rate adjustment and data augmentation technology, it optimizes the image data in the road disease dataset to obtain preprocessed image data. The model building module is used to build a YOLOv11 detection model, in which a sub-pixel-level edge enhancement module and a wavelet transform feature extraction module are embedded in the YOLOv11 framework. The model training and optimization module is used to adopt a phased training strategy. First, the YOLOv11 detection model is pre-trained using a large natural image dataset in the public domain. Then, the YOLOv11 detection model is trained using the preprocessed image data obtained from step 2. The model evaluation and deployment module is used to evaluate the performance of the trained YOLOv11 detection model on the test set; and to perform defect detection on the road surface image under test using the trained YOLOv11 detection model.

8. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the YOLOv11 road defect detection method based on multi-scale feature enhancement as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program causes the computer to execute a YOLOv11 road defect detection method based on multi-scale feature enhancement as described in any one of claims 1 to 6.

Images