Research On Road Demage Detection Algorithm Based On Deep Learning

At present,with the rapid development of the economy and society of China,urbanization process is accelerating,road network scale is also expanding.However,roads are engineered structures with a certain service life,so it is particularly imperative to maintain them.Road damage must be detected and repaired in time to keep its performance and prolong its reconstruction time.Traditional image processing technology has been widely used in this field,but it has limitations such as low detection efficiencies and the result is greatly affected by the environment.Using deep learning theory as a starting point,this paper investigates in-depth deep learning-based methods for detecting road damage,the main work is as follows:1.Aiming at the lack of public datasets covering comprehensive damage types for road damage detection research in China nowadays,a dataset containing 8 common road damage types,8312 road damage images and 15827 road damage label boxes was produced.Aiming at the problem of less data samples,transfer learning method is used to train the model,and data enhancement technology is used to improve the imbalance of samples,single image and other problems in the dataset,so as to effectively improve the the generalization ability of deep learning models.2.A road damage detection method based on feature extraction is proposed,which optimizes the backbone of the Faster R-CNN algorithm.Firstly,the backbone of Faster R-CNN is replaced by ResNet50 with deeper network,and the FPN(Feature Pyramid Network)is fused with it to better extract features,so as to improve the detection accuracy of the model for road damage.Secondly,the network fused with MobileNet V3＿large and FPN is used as its backbone to improve the detection speed of the model on the basis of ensuring certain detection accuracy.3.A road damage detection method based on model pruning is proposed,which improves the YOLOv5s(v6.0)algorithm.First,the bounding box loss function of YOLOv5s(v6.0)is changed to SIoU＿Loss,second,the SimAM,a parameter-free soft attention mechanism,is added to its backbone to improve the efficiency of the network model in extracting training image information,last,the trained improved model is pruned using a regularized channel pruning strategy to reduce the model size and improve the speed of the model for road damage detection.