Road disease adaptive intelligent identification method based on deep learning and information perception
By constructing a complex background augmentation dataset and improving the deep learning model, the problems of anti-interference and lightweight deployment of road defect detection in harsh environments were solved, and high-precision road defect detection and real-time inspection were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176263A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of road defect identification technology, and in particular to an adaptive intelligent identification method for road defects based on deep learning and information perception. Background Technology
[0002] With the rapid development of transportation infrastructure, monitoring the health of road surfaces is crucial. Currently, road damage detection mainly relies on manual inspection or traditional machine vision methods. However, manual inspection is inefficient, risky, and highly subjective; traditional image processing algorithms based on threshold segmentation or edge detection are extremely sensitive to noise; and existing general-purpose object detection models based on deep learning (such as the YOLO series and Faster R-CNN) have serious shortcomings in complex real-world scenarios.
[0003] Background interference is severe: shadows of trees or buildings, water stains on the road, oil stains, etc. are easily misidentified as cracks or potholes.
[0004] Loss of subtle features: Traditional convolution operations are difficult to capture tiny cracks that are thin and have large curvature changes, and are very easy to miss in low light or foggy conditions.
[0005] High computational overhead and insufficient feature extraction: Existing models struggle to balance global semantics and local details when processing high-resolution images, and their large number of parameters makes them difficult to deploy on resource-constrained platforms such as automotive embedded devices.
[0006] To address the aforementioned issues, we propose an adaptive intelligent identification method for road defects based on deep learning and information perception. Summary of the Invention
[0007] In view of this, the purpose of this invention is to propose an adaptive intelligent identification method for road defects based on deep learning and information perception, so as to solve the problems of poor anti-interference ability of existing technologies in harsh environments, easy to miss minor damage, and difficulty in lightweight deployment of models.
[0008] To achieve the above objectives, this invention provides an adaptive intelligent identification method for road defects based on deep learning and information perception, comprising the following steps:
[0009] S1. Construct a complex background enhancement dataset: Obtain basic road damage image data, and use the style-aware state space model SaMam to transfer the complex environment style of the real world to the basic road damage images to generate an enhanced dataset containing twilight or fog mode.
[0010] S2. Construct a deep learning detection model based on the improved RT-DETR: Using RT-DETR as the basic architecture, introduce the C2f-FMB module based on feature mixing blocks into its backbone network, and introduce the DeformableMHSA-AIFI module based on deformable multi-head attention into its neck network.
[0011] S3. Model Training and Optimization: The deep learning detection model is trained using augmented datasets, and the model parameters are optimized using a loss function combining GIoULoss and Focal Loss.
[0012] S4. Online Real-time Detection Output: Input the image to be detected into the trained model. After filtering out background interference through forward inference, the model outputs the type, confidence level, and location coordinates of the road defects.
[0013] Preferably, in step S1, the style-aware state space model SaMam is a deep neural network that includes an input mapping layer (PatchEmbed), a visual state space module (VSSM), and a style-aware visual state space module (SAVSSM).
[0014] The style transfer process of the style-aware state-space model SaMam includes:
[0015] S1.1 Input the road crack content image and target style images Images (such as those taken at dusk or in fog) are divided into non-overlapping image patches, which are then mapped to the feature space through an input mapping layer (PatchEmbed) to obtain content features. and style characteristics ;
[0016] S1.2 Utilize the Visual State Space Module (VSSM) to globally model the features of the serialized image patches, thereby simulating the visual style of complex environments such as dusk and fog without affecting the structural integrity of road damage;
[0017]
[0018] in, Extract information such as road structure and damage morphology. Then, model the light distribution, contrast attenuation, and global appearance statistics under twilight or foggy conditions. Indicates style information of the image. This represents the content information of the image;
[0019] S1.3 During the decoding stage, style information is injected into the content information through the Style-Aware Visual State Space Module (SAVSSM), as shown in the following formula:
[0020]
[0021] in, Represented as channel scale modulation coefficients generated from style information, used to control the amplification or suppression of content information in different channels; Represented as a bias term generated from style information, used to introduce style-related appearance changes; This represents an element-wise multiplication operation. To represent road crack images with visual characteristics of dusk or fog, complex environmental appearance changes are introduced while keeping the crack geometry unchanged.
[0022] S1.4 Perform batch style transfer processing on the training set images to generate twilight mode and fog mode images, and merge them with the original images to construct the final complex scene road damage dataset.
[0023] Preferably, in step S2, the C2f-FMB module comprises N cascaded FMB blocks, each FMB block containing a spatial mixing sublayer and a channel mixing layer, and all using residual connections to ensure gradient propagation. Let the nth... The input of each FMB block is The output is The calculation logic is as follows:
[0024]
[0025]
[0026] in, This refers to the multi-head feature attention module, which is responsible for extracting long-range spatial dependencies. This is a parallel convolutional feedforward module, responsible for channel feature fusion.
[0027] Preferably, the spatial mixing sublayer adopts Modules are used to enhance the characteristic response of road damage areas under complex background conditions;
[0028] The module implements feature enhancement through two parallel branches;
[0029] Spatial attention weight modeling branch: joint downsampling spatial context features With channel-level global variance features Generate adaptive attention weights ;
[0030]
[0031] in, and For learnable parameters, The activation function is non-linear; the generated attention weights Used for element-wise modulation of the original features;
[0032] Local detail feature enhancement branch: The input features are modeled using local detail enhancement operators, and residual fusion is performed between the input features and the modulated features. Finally, the enhanced feature representation is output through a linear mapping. :
[0033] .
[0034] Preferably, the channel mixing layer adopts The module reduces computational redundancy and only performs local updates on intermediate features;
[0035] The The module introduces the concept of partial convolution and adopts an inverted residual structure;
[0036] Intermediate features along the channel dimension Divided into subsets that participate in convolution operations and directly preserving subsets through identity mapping The formula is as follows;
[0037] ;
[0038] The inverse residual structure includes sequentially executed dimensionality-expanding convolution, depthwise convolution, and dimensionality-reducing convolution operations, used to... Update;
[0039] Update the feature subset and The concatenation is performed along the channel dimension, and the calculation formula is as follows:
[0040]
[0041] in, This represents a nonlinear activation function (such as SiLU or ReLU). This represents a convolution operation that includes an inverse residual structure. This indicates a channel splicing operation.
[0042] Preferably, in step S2, the DeformableMHSA-AIFI module captures road distress features through the following process:
[0043] Prediction phase: A lightweight convolutional predictor is used to map the input features and predict each reference point on the feature map. Spatial offset and attention weight ;
[0044] Sampling phase: using the bilinear sampling operator Extract feature values at the offset position. The attention output is obtained by weighting and aggregating the attention values. ;
[0045] ;
[0046] Fusion output stage: With residual input features Element-wise addition is performed followed by layer normalization, and then the mixture is passed through a feedforward network. Enhance feature representation capabilities to obtain the module's final output. The formula is as follows:
[0047] .
[0048] Preferably, the model training process in step S3 specifically includes:
[0049] Training parameter configuration: Set the input image resolution to 640×640 and use the AdamW optimizer for training;
[0050] Loss function construction: A combination of GIoU Loss and Focal Loss is adopted. GIoU Loss is used to improve positioning accuracy, and Focal Loss is used to solve the imbalance problem where the road background is much larger than the disease sample.
[0051] Weight preservation strategy: Update network weights through backpropagation algorithm, and save the optimal model weights when the average accuracy mAP on the validation set no longer improves.
[0052] Preferably, the online real-time detection output of step S4 specifically includes:
[0053] Real-time image input: Real-time road surface images are collected using inspection equipment and fed into the trained improved model;
[0054] Feature extraction and interference filtering: During the forward inference process, the model filters out road shadows and complex background interference through the cross-stage local fusion module and the intra-scale feature interaction module, and extracts pure crack features;
[0055] Detection result output: The detection head outputs the type, confidence level, and precise location coordinates of road defects, and performs visual annotation on the original image; the type includes at least one of longitudinal cracks, transverse cracks, network cracks, potholes, and pavement repair.
[0056] A road defect detection model that performs the above-described method, the model using an improved RT-DETR as its architecture, including a C2f-FMB module based on feature fusion blocks in its backbone network and a DeformableMHSA-AIFI module based on deformable multi-head attention in its neck network.
[0057] A road defect detection system includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the steps of the above method.
[0058] The beneficial effects of this invention are as follows:
[0059] I. This invention addresses challenging environments such as fog and dusk by expanding a highly realistic complex scene dataset using the style-aware state space model SaMam at the data level. At the model level, it introduces Multi-Head Feature Attention (SMFA) and Deformable Attention (MHSA-AIFI). Through dynamic adjustment of attention weights, the model adaptively suppresses high-frequency background noise such as tree shadows and road surface water stains, significantly reducing the false detection rate.
[0060] Second, for fine cracks and irregular pits, the C2f-FMB module designed in this invention expands the receptive field through large-kernel depth convolution to capture the global connectivity of the cracks. Simultaneously, the deformable sampling mechanism in the DeformableMHSA-AIFI module allows sampling points to break through the fixed grid and deform towards the crack edges or pit centers. This significantly improves the model's ability to extract fine-grained features at low contrast, effectively reducing missed detections.
[0061] Third, this invention abandons the computationally intensive global similarity calculation of Query-Key in traditional Transformers, and adopts partial convolution and a lightweight attention mechanism based on locally deformable aggregation. While improving detection performance, it significantly reduces computational complexity (FLOPs) and memory usage, enabling the model to be easily deployed on resource-constrained platforms such as in-vehicle embedded devices and drones, meeting real-time inspection requirements. Attached Figure Description
[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0063] Figure 1 This is a flowchart of the steps of the present invention;
[0064] Figure 2 This is a schematic diagram of the style-aware state-space model of the present invention;
[0065] Figure 3 This is a schematic diagram illustrating the overall architecture of the improved RT-DETR detection model of this invention.
[0066] Figure 4 This is a schematic diagram of the cross-stage local fusion module of the present invention;
[0067] Figure 5 This is a schematic diagram of the scale-based feature interaction module of the present invention. Detailed Implementation
[0068] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0069] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this invention should have the ordinary meaning understood by those skilled in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0070] like Figures 1-5 As shown, the adaptive intelligent identification method for road defects based on deep learning and information perception includes the following steps:
[0071] S1. Construct a complex background enhancement dataset: Obtain basic road damage image data and select a set of road damage images collected from multiple angles and sources (in this embodiment, the RDD2022-China dataset is selected). The dataset includes aerial images taken by UAVs and images taken at eye level by vehicle-mounted cameras, covering five typical damage targets: longitudinal cracks (D00), transverse cracks (D10), network cracks (D20), potholes (D40), and pavement repair.
[0072] To prevent cross-set information leakage during data augmentation and ensure the objectivity of model evaluation, the base image set is first divided into training, validation, and test sets according to a predetermined ratio (e.g., 8:1:1). Subsequent data augmentation and style transfer operations are strictly performed independently within each subset.
[0073] Using the style-aware state space model SaMam, the styles of complex real-world environments are transferred to basic road damage images to generate augmented datasets containing twilight or foggy modes.
[0074] Construct a style-aware state-space model (SaMam) based on a deep neural network, such as Figure 2 As shown, the model includes an input mapping layer (PatchEmbed), a visual state space module (VSSM), and a style-aware visual state space module (SAVSSM).
[0075] The style transfer process of the style-aware state-space model (SaMam) includes:
[0076] S1.1 Input the road crack content image and target style images Images (such as those taken at dusk or in fog) are divided into non-overlapping image patches, which are then mapped to the feature space through an input mapping layer (PatchEmbed) to obtain content features. and style characteristics ;
[0077] S1.2 Utilize the Visual State Space Module (VSSM) to globally model the features of the serialized image patches, thereby simulating the visual style of complex environments such as dusk and fog without affecting the structural integrity of road damage;
[0078]
[0079] in, Extract information such as road structure and damage morphology. Then, model the light distribution, contrast attenuation, and global appearance statistics under twilight or foggy conditions. Indicates style information of the image. This represents the content information of the image;
[0080] S1.3 During the decoding stage, style information is injected into the content information through the Style-Aware Visual State Space Module (SAVSSM), as shown in the following formula:
[0081]
[0082] in, Represented as channel scale modulation coefficients generated from style information, used to control the amplification or suppression of content information in different channels; Represented as a bias term generated from style information, used to introduce style-related appearance changes; This represents an element-wise multiplication operation. To represent road crack images with visual characteristics of dusk or fog, complex environmental appearance changes are introduced while keeping the crack geometry unchanged.
[0083] S1.4. Batch style transfer processing was performed on the training set images to generate twilight and foggy mode images, which were then merged with the original images to construct the final complex scene road damage dataset (RDD2022_ChinaST). This dataset contains 10,506 training images, 1,311 validation images, and 1,317 test images. This dataset, by simulating harsh real-world imaging environments, bridges the domain gap between laboratory data and practical application scenarios.
[0084] S2. Construct a deep learning detection model based on the improved RT-DETR, such as... Figure 3 As shown, this embodiment uses RT-DETR as its basic architecture and makes targeted improvements to its backbone and neck networks.
[0085] A cross-stage local fusion module (C2f-FMB module) based on feature mixing blocks is introduced into its backbone network to address the problems of limited receptive field and feature coupling in traditional convolutions, such as... Figure 4 As shown, the traditional C2f module is replaced with a C2f-FMB module, which contains N cascaded feature mixing blocks (FMB blocks). Each FMB block contains a spatial mixing sublayer and a channel mixing layer, and all use residual connections to ensure gradient propagation. Let the nth... The input of each FMB block is The output is The calculation logic is as follows:
[0086]
[0087]
[0088] in, This refers to the multi-head feature attention module, which is responsible for extracting long-range spatial dependencies. This is a parallel convolutional feedforward module, responsible for channel feature fusion.
[0089] The spatial hybrid sublayer adopts Modules are used to enhance the characteristic response of road damage areas under complex background conditions;
[0090] The module implements feature enhancement through two parallel branches;
[0091] Spatial attention weight modeling branch: joint downsampling spatial context features With channel-level global variance features Generate adaptive attention weights ;
[0092]
[0093] in, and For learnable parameters, The activation function is non-linear; the generated attention weights Used for element-wise modulation of the original features;
[0094] Local detail feature enhancement branch: The input features are modeled using local detail enhancement operators, and residual fusion is performed between the input features and the modulated features. Finally, the enhanced feature representation is output through a linear mapping. :
[0095] .
[0096] The channel mixing layer adopts The module reduces computational redundancy and only performs local updates on intermediate features;
[0097] The The module introduces the concept of partial convolution and adopts an inverted residual structure;
[0098] Intermediate features along the channel dimension Divided into subsets that participate in convolution operations and directly preserving subsets through identity mapping The formula is as follows;
[0099] ;
[0100] The inverse residual structure includes sequentially executed dimensionality-expanding convolution, depthwise convolution, and dimensionality-reducing convolution operations, used to... Update;
[0101] Update the feature subset and The concatenation is performed along the channel dimension, and the calculation formula is as follows:
[0102]
[0103] in, This represents a nonlinear activation function (such as SiLU or ReLU). This represents a convolution operation that includes an inverse residual structure. This indicates a channel splicing operation.
[0104] A deformable multi-head attention-based intra-scale feature interaction module (DeformableMHSA-AIFI module) is introduced into the neck network; this addresses the irregular and weakly textured characteristics of cracks in harsh environments, such as... Figure 5 As shown, this invention improves the feature interaction stage. It abandons the computationally intensive global Query-Key similarity calculation. In step S2, the DeformableMHSA-AIFI module captures road defect features through the following process:
[0105] Prediction phase: A lightweight convolutional predictor is used to map the input features and predict each reference point on the feature map. Spatial offset and attention weight ;
[0106] Sampling phase: using the bilinear sampling operator Extract feature values at the offset position. The attention output is obtained by weighting and aggregating the attention values. ;
[0107] ;
[0108] Fusion output stage: With residual input features Element-wise addition is performed followed by layer normalization, and then the mixture is passed through a feedforward network. Enhance feature representation capabilities to obtain the module's final output. The formula is as follows:
[0109] .
[0110] S3. Model Training and Optimization: The deep learning detection model is trained using the augmented dataset (RDD2022-ChinaST). The network weights are updated by combining GIoU Loss (to improve the localization accuracy of the predicted bounding boxes) and Focal Loss (to solve the extreme sample imbalance problem where the road background is much larger than the defects). When the mAP (mean accuracy) of the validation set no longer improves, the optimal model weights are saved to complete the training.
[0111] The model training process in step S3 specifically includes:
[0112] Training parameter configuration: Set the input image resolution to 640×640 and use the AdamW optimizer for training;
[0113] Loss function construction: A combination of GIoU Loss and Focal Loss is adopted. GIoU Loss is used to improve positioning accuracy, and Focal Loss is used to solve the imbalance problem where the road background is much larger than the disease sample.
[0114] Weight preservation strategy: Update network weights through backpropagation algorithm, and save the optimal model weights when the average accuracy mAP on the validation set no longer improves.
[0115] S4. Online Real-time Detection Output: Input the image to be detected into the trained model. After filtering out background interference through forward inference, the model outputs the type, confidence level, and location coordinates of the road defects.
[0116] The online real-time detection output of step S4 specifically includes:
[0117] Real-time image input: Real-time road surface images are collected using inspection equipment and fed into the trained improved model;
[0118] Feature extraction and interference filtering: During the forward inference process, the model filters out road shadows and complex background interference through the cross-stage local fusion module and the intra-scale feature interaction module, and extracts pure crack features;
[0119] Detection result output: The detection head outputs the type, confidence level, and precise location coordinates of road defects, and performs visual annotation on the original image; the type includes at least one of longitudinal cracks, transverse cracks, network cracks, potholes, and pavement repair.
[0120] A road defect detection model that performs the above-described method, the model using an improved RT-DETR as its architecture, including a C2f-FMB module based on feature fusion blocks in its backbone network and a DeformableMHSA-AIFI module based on deformable multi-head attention in its neck network.
[0121] A road defect detection system includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the steps of the above method.
[0122] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.
[0123] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A road defect adaptive intelligent identification method based on deep learning and information perception, characterized in that, Includes the following steps: S1. Construct a complex background enhancement dataset: Obtain basic road damage image data, and use the style-aware state space model SaMam to transfer the complex environment style of the real world to the basic road damage images to generate an enhanced dataset containing twilight or fog mode. S2. Construct a deep learning detection model based on the improved RT-DETR: Using RT-DETR as the basic architecture, introduce the C2f-FMB module based on feature mixing blocks into its backbone network, and introduce the DeformableMHSA-AIFI module based on deformable multi-head attention into its neck network. S3. Model Training and Optimization: The deep learning detection model is trained using augmented datasets, and the model parameters are optimized using a loss function combining GIoULoss and Focal Loss. S4. Online Real-time Detection Output: Input the image to be detected into the trained model. After filtering out background interference through forward inference, the model outputs the type, confidence level, and location coordinates of the road defects.
2. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 1, characterized in that, In step S1, the style-aware state space model SaMam is a deep neural network that includes an input mapping layer, a visual state space module, and a style-aware visual state space module. The style transfer process of the style-aware state-space model SaMam includes: S1.1 Input the road crack content image and target style images The images are divided into non-overlapping patches and mapped to the feature space through an input mapping layer to obtain content features. and style characteristics ; S1.2 Utilize the visual state space module to perform global modeling of the features of the serialized image blocks, thereby simulating the visual style of complex environments such as dusk and fog without affecting the structural integrity of road damage; ， in, Extract information such as road structure and damage morphology. Then, model the light distribution, contrast attenuation, and global appearance statistics under twilight or foggy conditions. Indicates style information of the image. This represents the content information of the image; S1.3 During the decoding stage, style information is injected into the content information through the style-aware visual state space module, as shown in the following formula: ， in, Represented as channel scale modulation coefficients generated from style information, used to control the amplification or suppression of content information in different channels; Represented as a bias term generated from style information, used to introduce style-related appearance changes; This represents an element-wise multiplication operation. To represent road crack images with visual characteristics of dusk or fog, complex environmental appearance changes are introduced while keeping the crack geometry unchanged. S1.4 Perform batch style transfer processing on the training set images to generate twilight mode and fog mode images, and merge them with the original images to construct the final complex scene road damage dataset.
3. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 1, characterized in that, In step S2, the C2f-FMB module contains N cascaded FMB blocks. Each FMB block contains a spatial mixing sublayer and a channel mixing layer, and all use residual connections to ensure gradient propagation. Let the nth... The input of each FMB block is The output is The calculation logic is as follows: ， ， in, This refers to the multi-head feature attention module, which is responsible for extracting long-range spatial dependencies. This is a parallel convolutional feedforward module, responsible for channel feature fusion.
4. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 3, characterized in that, The spatial hybrid sublayer adopts Modules are used to enhance the characteristic response of road damage areas under complex background conditions; The module implements feature enhancement through two parallel branches; Spatial attention weight modeling branch: joint downsampling spatial context features With channel-level global variance features Generate adaptive attention weights ; ， in, and For learnable parameters, The activation function is non-linear; the generated attention weights Used for element-wise modulation of the original features; Local detail feature enhancement branch: The input features are modeled using local detail enhancement operators, and residual fusion is performed between the input features and the modulated features. Finally, the enhanced feature representation is output through a linear mapping. : 。 5. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 3, characterized in that, The channel mixing layer adopts The module reduces computational redundancy and only performs local updates on intermediate features; The The module introduces partial convolutional concepts and employs an inverted residual structure; Intermediate features along the channel dimension Divided into subsets that participate in convolution operations and directly preserving subsets through identity mapping The formula is as follows; ； The inverse residual structure includes sequentially executed dimensionality-expanding convolution, depthwise convolution, and dimensionality-reducing convolution operations, used to... Update; Update the feature subset and The concatenation is performed along the channel dimension, and the calculation formula is as follows: ， in, Represents a non-linear activation function. This represents a convolution operation that includes an inverse residual structure. This indicates a channel splicing operation.
6. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 1, characterized in that, In step S2, the DeformableMHSA-AIFI module captures road distress features through the following process: Prediction phase: A lightweight convolutional predictor is used to map the input features and predict each reference point on the feature map. Spatial offset and attention weight ; Sampling phase: using the bilinear sampling operator Extract feature values at the offset position. The attention output is obtained by weighting and aggregating the attention values. ; ； Fusion output stage: With residual input features Element-wise addition is performed followed by layer normalization, and then the mixture is passed through a feedforward network. Enhance feature representation capabilities to obtain the module's final output. The formula is as follows: 。 7. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 1, characterized in that, The model training process in step S3 specifically includes: Training parameter configuration: Set the input image resolution to 640×640 and use the AdamW optimizer for training; Loss function construction: A combination of GIoU Loss and Focal Loss is adopted. GIoU Loss is used to improve positioning accuracy, and Focal Loss is used to solve the imbalance problem where the road background is much larger than the disease sample. Weight preservation strategy: Update network weights through backpropagation algorithm, and save the optimal model weights when the average accuracy mAP on the validation set no longer improves.
8. The adaptive intelligent identification method for road defects based on deep learning and information perception according to claim 1, characterized in that, The online real-time detection output of step S4 specifically includes: Real-time image input: Real-time road surface images are collected using inspection equipment and fed into the trained improved model; Feature extraction and interference filtering: During the forward inference process, the model filters out road shadows and complex background interference through the cross-stage local fusion module and the intra-scale feature interaction module, and extracts pure crack features; Detection result output: The detection head outputs the type, confidence level, and precise location coordinates of road defects, and performs visual annotation on the original image; the type includes at least one of longitudinal cracks, transverse cracks, network cracks, potholes, and pavement repair.
9. A road defect detection model, characterized in that, The model performs the method as described in any one of claims 1-8, the model using an improved RT-DETR as its architecture, including a C2f-FMB module based on feature mixing blocks in its backbone network, and a DeformableMHSA-AIFI module based on deformable multi-head attention in its neck network.
10. A road defect detection system, characterized in that, It includes a processor and a memory, wherein when the processor executes a computer program stored in the memory, it implements the steps of the method as described in any one of claims 1-8.