Pothole road detection and identification based on improved DeepLab V3+

Abstract

This paper proposes an improved DeepLab V3+ model based on attention mechanism, which can accurately locate road potholes and effectively extract multi-scale semantic information. The improved model uses the combination of Canny edge extraction algorithm and deep learning method to input rough edges as additional semantic information into the model, and introduces channel attention mechanism and separable cavity convolutional neural network to optimize feature extraction and model size. The effectiveness of the model is proved by training the model on the open source data set and testing its performance in multiple dimensions. The trained model is used to deduce the image on the test set, and the binary mask of the road pothole is obtained, and the refined edge is obtained by edge extraction algorithm. For the second problem, the pothole part of the semantic segmentation map is counted by pixel points, and the proportion of pothole area in the whole image area is obtained, and the result is written into the file. Finally, the advantages and disadvantages of the model are evaluated, and the application of the model in other fields is discussed.