Research On Road Damage Detection Model Based On Deep Learning

The deterioration of roads poses a threat to road safety and can create obstacles for driving safety,personal safety,and urban construction when severe.Timely detection of road damage is crucial for road hazard prevention and subsequent road maintenance.Therefore,efficient and intelligent detection methods have become an urgent need for road damage detection.In recent years,computer vision techniques based on deep learning have become a new direction for object detection.In this study,we proposed a more accurate and lower parameter YOLO-RDD algorithm based on the one-stage object detection algorithm YOLOv5 m,using the open-source dataset GRDDC2022 as the basis.The algorithm was deployed on both PC and Web platforms to meet the needs of practical application scenarios.The main research results and related conclusions are as follows:(1)Based on the Chinese regional data sources in the GRDDC2022 dataset,we performed single-sample data augmentation and multi-sample data augmentation.The single-sample data augmentation method included geometric transformation,color transformation,noise transformation,and local erasure.The multi-sample data augmentation method included mixed images,label smoothing,and Gridmask.To address the problem of the significant differences in the number of damage categories in the original dataset,with fewer samples of alligator cracks(D20)and potholes(D40),we used the excellent performance of the FAST＿GAN model to generate 200 D20 and D40 damage instances and used Poisson image blending to mix the generated damage instance images with the road background images to produce new road damage images.After completing the labeling in Labelimg,they were added as part of the training dataset,completing the construction of the new road damage dataset,with 7,004 training images and 876 validation images.The experimental results showed that the model trained using the augmented dataset had higher detection accuracy.Before using data augmentation,the MAP@0.5 of the YOLOv5 model was only 0.732,while after using the augmented dataset,the model’s MAP@0.5 reached 0.878,an increase of 14.6%.Moreover,the detection accuracy of each category also improved significantly,indicating that data augmentation operations are crucial for improving model accuracy.(2)The present study proposes a lightweight detection model,named YOLO-RDD,which is based on the first-stage object detection model YOLOv5.The Kmeans++ algorithm is employed for pre-selecting box clustering on the dataset to select more suitable anchors.To improve traditional Conv and Bottleneck modules,Ghostnet is introduced into the network structure.Sim Conv is used to redesign the spatial pyramid pooling layer Sim SPPF,and attention mechanism is introduced to the network’s neck part to enhance the feature extraction ability for small objects.CARAFE is used for the feature map sampling,which shows superior performance.Finally,the multi-scale feature fusion network Bi FPN borrowed from Efficient Det is utilized to improve the feature map fusion method in the neck part of the model.The angle loss is also considered in the bounding box loss function to speed up the model training.The model is trained on the enhanced dataset and compared with the current mainstream 5-class detection models.The experimental results demonstrate that the YOLO-RDD model performs very well in terms of detection accuracy and model parameter quantity,achieving the best results in both MAP@0.5and F1-Score indicators.The YOLO-RDD model surpasses larger YOLOv5 L models and new versions of YOLOv6,YOLOv7,and YOLOv8 algorithms in terms of detection performance.Moreover,the YOLO-RDD model is lightweight,having the least number of parameters among all compared algorithms,with a total parameter count of 16726284,which is 4138773 less than YOLOv5.The final training weight of the YOLO-RDD model is only 32.4Mb,which is the smallest among all compared algorithms.The detection speed of the YOLO-RDD model reaches30.9FPS,which meets the standard of real-time detection.(3)In order to meet the application needs of different users for road damage detection algorithms,this study deploys the YOLO-RDD algorithm to PC and Web ends,achieving realtime detection of road images and video stream data,and verifying the practicality and effectiveness of the proposed algorithm.