Road Damage Detection Based On Deep Learning

In recent years,with the rapid development of road traffic,road surface damage is getting worse,which not only greatly increases road surface maintenance tasks,but also may cause excessive wear and tear of vehicles,increase the probability of traffic accidents,cause unnecessary personal and property losses,and affect traffic efficiency and economic and social development.Therefore,the demand for road damage detection is increasing and the requirements are getting higher.Accurate and efficient road damage detection is a very urgent practical need.In the real scene,road condition is complex and diverse,and is easily affected by various environmental factors such as light and shadow.Traditional image processing methods cannot detect road damage effectively.Due to the particularity of the road damage dataset,damage targets are mostly in the shape of thin strips,occupy few pixels,and contain insufficient effective features.Using a lightweight architecture can lead to insufficient accuracy,while using a complex model cannot meet the requirements of real-time detection.Therefore,it is necessary to study and establish a more advanced road damage detection model,increasing the extraction of effective features,and achieving high-precision and fast road damage detection.Considering the real-time and accuracy requirements of the road damage detection task,in this work,we propose a road damage detection model YOLO-AC based on deep learning.We use YOLOv5 s as the basic structure of the network.By adding a deeper feature extraction layer in the Backbone network,deeper semantic information can be obtained.At the same time,by adding the corresponding feature detection layer and cooperating with the Anchor mechanism to re-subdivide the scale,then dense targets can be assigned to different detection layers,and further realize the multi-scale divide-and-conquer detection strategy.the coordinate attention module(CA)is introduced,so that it can adaptively filter the extracted important information,ignore irrelevant information,help the network integrate important features,locate the target more accurately,and improve detection accuracy;We employ Bi FPN as the feature fusion structure in the Neck network,which weighted two-way fusion features.The Context Augmentation Module(CAM)is introduced to supplement multi-scale context information and strengthen the learning of the network features;The anchor box generated by the K-means clustering algorithm is more suitable for the current dataset,making the network easier to learn,and the training process more stable,and helping the model to converge quickly;In addition,according to the characteristics of the road damage dataset,the orientation information of cracks is introduced.Based on EIOU,D-EIOU loss is proposed as localization loss.By introducing orientation information,the model can better learn the essential features of cracks in different directions.At last,we conduct experiments on the RDD2020 dataset.The experimental results show that the method proposed can effectively improve detection accuracy.At the same time,it meets the detection requirements of good real-time performance,and the lightweight model is easier to be deployed on mobile devices.