CN-122023226-A - Road defect detection method and model based on lightweight enhanced feature fusion

CN122023226ACN 122023226 ACN122023226 ACN 122023226ACN-122023226-A

Abstract

The invention relates to the field of computer vision and intelligent traffic, and discloses a road defect detection method and model based on lightweight enhanced feature fusion. The method is based on the YOLO11 architecture, including backbone network, neck network and head network. A simplified heavy parameter batch normalization module (NRepBN) is adopted in the backbone network, and a residual normalization path is introduced in a training stage to improve stability. And an enhanced relevant feature fusion module (ECFM) is introduced into the neck network, and the fine crack features are effectively reserved through grouping feature focusing and multi-layer feature reconstruction, so that the multi-scale correlation is enhanced. The training process combines the self-adaptive threshold focusing loss function (ATFL), dynamically adjusts the focusing parameters according to the prediction confidence, and relieves the performance degradation caused by category imbalance. The method and the model can improve the detection precision of small targets such as micro cracks, pits and the like in road defect image detection, simultaneously keep the light-weight characteristic, and are suitable for road defect detection of various scenes and real-time deployment on edge equipment.

Inventors

ZHU CHUCHU
LI JI
CHEN YUXIAO
ZHANG ZHIPENG
Deng Zhaotian

Assignees

桂林电子科技大学
广西联睿智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251010

Claims (7)

1. A road defect detection method and model based on lightweight enhanced feature fusion are characterized by comprising the following steps: s1, acquiring a road defect image, and dividing the data set into a training set, a verification set and a test set; S2, constructing an ERAN-YOLO network model based on a YOLO11 architecture; s3, training the ERAN-YOLO network model in the S2 by utilizing the training set in the S1 to obtain a road defect detection model after training; and S4, carrying out road defect detection verification in a verification set by adopting the model trained in the step S3, and carrying out final detection test in a test set to output the category and position information of the road defect.
2. The road defect detection method and model based on lightweight enhanced feature fusion according to claim 1, wherein the specific steps of step S1 are as follows: S11, obtaining a road defect image in an RDD2022 data set, wherein the data set comprises road defect images such as longitudinal cracks, transverse cracks, network cracks and pits; S12, following official training set division rules, dividing partial samples from the training set to serve as a verification set by using a random sampling method.
3. The road defect detection method and model based on lightweight enhanced feature fusion according to claim 1, wherein the specific steps of step S2 are as follows: s21, an ERAN-YOLO network model is improved based on YOLO11, and the whole architecture is consistent with that of YOLO11 and sequentially comprises a backbone network, a neck network and a head network; s22, setting a simplified heavy parameter batch normalization module (NRepBN) in a backbone network, introducing a residual normalization path in a training stage to improve training stability and fusing with a previous convolution layer in an reasoning stage to reduce calculation overhead; S23, setting an enhancement relevant feature fusion module (ECFM) in a neck network, and realizing the reservation of fine crack features and multi-scale relevance enhancement through grouping feature focusing and multi-layer feature reconstruction; And S24, adopting an adaptive threshold focusing loss function (ATFL) in the training process, and dynamically adjusting the focusing parameters according to the prediction confidence level so as to relieve the detection performance reduction caused by the class imbalance.
4. The road defect detection method and the model based on lightweight enhanced feature fusion according to claim 3, wherein the specific steps of the network model NRepBN module in step S22 are as follows: s221, introducing a residual error normalization path after the output of the convolution layer, removing a gradual transition mechanism, and reserving a core normalization residual path, wherein the residual error normalization path is defined by a formula (1); NRepBN(x)=BN(x)+α·x (1) Where x is the input feature tensor, BN (x) is the batch normalization result, and α is the learnable parameter for balancing the normalized branch with the residual branch. S222, fusing the batch normalization parameters with the convolution weight of the previous layer in the reasoning stage so as to reduce the calculation overhead and improve the reasoning speed.
5. The road defect detection method and model based on lightweight enhanced feature fusion according to claim 3, wherein the specific steps of the network model ECFM module in step S23 are as follows: S231, setting a channel matching convolution layer for unifying the channel number of the input features, wherein the parameter calculation mode is defined by a formula (2); Params match ＝1×1×C 1 ×C 2 (2) Wherein, C 1 is the number of input channels, and C 2 is the number of output channels. S232, setting a space guidance fusion unit for generating a global attention map through lightweight convolution, and weighting important areas in the feature fusion process, wherein the parameter calculation mode is defined by a formula (3); Params global ＝3×3×C 2 +1×1×C 2 (3) S233, setting a grouping feature focusing unit for enhancing the correlation between grouping features and reducing the quantity of convolution parameters, wherein the calculation mode is defined by a formula (4), and the parameter calculation mode is defined by a formula (5); wherein P gi is the ith feature subgroup, U represents the set and operates, P i and P i-1 are the feature maps of the current layer and the previous layer respectively, Characteristic splicing is represented, mean and std are respectively the mean value and standard deviation of the spliced characteristic, P f is the normalized grouping characteristic output, C 2 is the output channel number, and G is the grouping number. S234, setting a multi-layer feature reconstruction unit for reconstructing small target features and reducing the loss of the small target features caused by deep convolution, wherein the calculation mode is defined by a formula (6); P m ＝Softmax(T(A(P low )))P new (6) and S235, setting an adjustable output control convolution layer for adjusting the number of output channels to adapt to the characteristic pyramid structure of the YOLO11, wherein the characteristic pyramid structure is defined by a formula (7). Params down ＝1×1×C 2 ×C 3 (7) Wherein, C 2 is the number of input channels, and C 3 is the number of output channels.
6. The road defect detection method and model based on lightweight enhanced feature fusion according to claim 3, wherein the specific steps of the loss function adopted by the network model training in step S24 are as follows: S241: ATFL adopts a confidence level smoothing strategy based on the prediction probability p t , and is defined by a formula (8); Where p t denotes the current prediction probability, For the historical smoothing of the confidence level, Is the updated confidence level. S242 based on Dynamically adjusting a focusing parameter gamma, defined by formula (9); where γ is a dynamically adjusted focusing factor. S243, re-weighting the loss value by adopting different tempering factors in different confidence intervals according to the size of the prediction probability so as to enhance the learning ability of the samples difficult to classify, wherein the learning ability is defined by a formula (10).
7. A road defect detection device based on lightweight enhanced feature fusion, characterized in that the road defect detection method and model based on lightweight enhanced feature fusion as claimed in claim 1 or 3 are implemented, comprising: the road defect image acquisition module is used for acquiring a road defect image which is divided into a training set, a verification set and a test set; the model training and verifying module is used for training the road defect detection model fused with the lightweight enhanced features based on the training set and verifying and testing the lightweight enhanced features on the verifying set and the testing set; the detection result output module is used for inputting the road defect image to be detected into the road defect detection model after training, and the type and the position information of the road defect in the image.

Description

Road defect detection method and model based on lightweight enhanced feature fusion Technical Field The invention relates to the technical field of computer vision and road detection, in particular to a road defect detection method and model based on lightweight enhanced feature fusion. Background Roads are important infrastructure for urban traffic and socioeconomic development, and the operational safety and life of roads are closely related to the level of road maintenance. The road defect detection has important practical significance for guaranteeing driving safety, reducing traffic accident rate, prolonging the service life of a road and reducing maintenance cost. The rapid identification and positioning of surface defects such as cracks and pits are one of the focus of attention in road surface defect detection. In a trunk road, the defects on the road surface can not be found and treated in time, so that potential safety hazards can be generated, and serious traffic accidents can be caused. Thus, road surface defect detection is an important ring of urban traffic safety and infrastructure maintenance. Conventional road defect detection mainly includes detection techniques based on image processing, machine learning, and three-dimensional imaging. The image processing includes threshold segmentation, edge detection, region growing and other modes, can effectively detect under a simple background, but has poor adaptability to complex background and variable illumination conditions. In road defect detection, machine learning improves adaptability to a certain extent by utilizing similarity among samples, but has limited characteristic expression capability, and is difficult to fully characterize complex road defect characteristics. The three-dimensional imaging technology has certain advantages in deep structural crack detection by acquiring the three-dimensional structural information of the pavement, but the three-dimensional imaging system has higher equipment cost and limited application range. In recent years, deep learning has rapidly progressed, and a target detection method based on a convolutional neural network has gradually become a mainstream technology for road defect detection. Among them, YOLO11 is an algorithmic model developed by the Ultralytics team for object detection and instance segmentation, which is a sophisticated, advanced (SOTA) model built on the basis of previous YOLO success, and introduces new functionality and improvements to further improve performance and flexibility. However, such methods still have shortcomings in terms of small target detection, class imbalance, real-time deployment of edge devices, and the like in a complex background. Therefore, how to provide a road defect detection method based on an improved YOLO11 model, which combines light weight, improved accuracy and enhanced robustness, so as to adapt to the application requirements of actual road inspection and intelligent traffic systems is a problem to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a road defect detection method and a model based on lightweight enhancement feature fusion, which are suitable for real-time deployment in road inspection and intelligent traffic systems by improving small target detection capability while keeping the weight of the model and relieving the problem of unbalanced road image categories so as to enhance the robustness of the model under a complex background. In order to solve the technical problems, the technical scheme adopted by the invention is that a road defect detection method and model based on an improved YOLO11 architecture are designed and trained, and the method specifically comprises the following steps: s1, acquiring a road defect image, wherein the road defect image is divided into a training set, a verification set and a test set; S2, model improvement, namely, according to a backbone network and a neck network of a technical demand improvement model, setting a simplified heavy parameter batch normalization module (NRepBN) in a plurality of convolution layers of the backbone network, improving training stability by introducing a residual normalization path in a training stage, and fusing batch normalization parameters with the convolution layers in a reasoning stage to reduce calculation expenditure; S3, training and verifying the model, namely training the lightweight enhanced feature fused road defect detection model by using a training set, and verifying the model on a verification set to obtain a trained road defect detection model, wherein the training process adopts a self-adaptive threshold focusing loss function (ATFL) to dynamically adjust the weight of a prediction sample, so that the improvement limit of detection performance caused by the class imbalance problem is relieved; and S4, defect detection, namely inputting a road defect image to be detected into a trained road defect detection mod