Asphalt Pavement Distress Recognition Based On Transformer

Pavement distress detection is a fundamental part of road operation and management,and information on distress will be used in future maintenance and planning.The cost of road maintenance will increase significantly as distress develops,so early detection of distress and planning for it at an early stage is key to road maintenance.In the past,the main method for detecting defects has been manual,which is not only time-consuming but also subjective.With the rapid development of deep learning in the field of computer vision,more and more fields are using deep learning methods to solve practical problems.Research on the application of deep learning models to pavement distress recognition will greatly improve the efficiency of detection and provide strong support for pavement maintenance decisions.Although there has been a lot of work applying deep learning models for distress detection,most of them suffer from a single type of detection target and are difficult to apply to practical detection,mainly because of the differences and similarities between distresses-distresses such as slender cracks are too distinct from other distresses,while at the same time,there are cracks that are easily confused with distress edges.In addition,the high cost of labelling data leads to a lack of data,and the distribution of data varies from region to region so the trained models tend to be overfitted.To address the problems of pavement disease detection,this paper designs and implements a self-supervised learning and Transformer based asphalt pavement disease detection To address the problems of pavement disease detection,this paper designs and implements a self-supervised learning and Transformer-based asphalt pavement disease identification system.The details of the research are as follows.Dataset creation.Two datasets are used in this paper:(1)a non-destructive discrimination dataset based on 336 high-precision grey-scale images of the Chongqing Airport Expressway.(2)The object detection dataset was produced based on 8321 images of Zhejiang Yongtaiwen pavement data and CQU-BPDD.Designed a model called RGB-Vi T to discriminate between pavement images with and without distress.This paper proposes a discriminative model based on Vi T and MAE pre-training for pavement library matching.In this paper,it is experimentally demonstrated that the method can effectively improve the performance and solve the performance drop of the model for out-of-distribution data.Designs a pavement distress identification method,GP-Detector.This paper improves and implements a detection method based on Swin-Transformer for pavement library prompts and demonstrates in this paper that it is effective in improving performance on the collected dataset and that the method can be easily plugged into other models.A web-based pavement disease recognition system is implemented.This paper develops an online pavement distress recognition system based on the proposed method.By uploading videos to be inspected,it is able to automatically detect and return the results,while providing a back-end inspection module that provides a complete interface for embedding in other web applications.This paper designs and implements an automatic detection module for pavement distress based on Transformer.The proposed method improves the robustness and accuracy of detection.The module can provide detection results for subsequent pavement maintenance and decisions,which has important application value.