Automatic Pavement Diseases Recognition Study Based On Deep Weakly Supervised Learning

With the development of China’s infrastructure,the road net is becoming surprisingly dense,and the maintenance of pavement is the key to a sustainable road net.To maintain a healthy transportation system,the recognition of pavement diseases is crucial for pavement maintenance and management.In the current road maintenance procedure,the recognition of road defects is basically done by the experienced worker through manual inspection.Such a procedure is not only inefficient,but the accuracy of recognition highly depends on the proficiency of the worker.In addition,it takes a lot of time to train such a skilled worker.Due to limited resources,different types of road diseases should be prioritized,and those serious type of road diseases should be given the highest priority.Therefore,it is imperative to conduct research on the automated recognition of pavement diseases.In this thesis,we present an end-to-end deep learning approach named Weakly Supervised Patch Label Inference Network with Image Pyramid(WSPLIN-IP)for recognizing pavement diseases,the main work of this thesis is as follows:(1)We propose a road diseases recognition method based on Res Net and Efficient Net.Using existing deep convolutional neural network models to solve the automatic pavement diseases recognition problem and compare the performance of several classical CNN models on the pavement diseases recognition task though extensive experimental analysis.The results show that Efficient Net method has better performance than other CNN models,and we also verify the advanced and superiority of deep learning methods over shallow learning methods.(2)To address the problem that CNN based pavement diseases recognition models cannot fully utilize the pixel information of high-resolution images,we present an end-to-end approach named Weakly Supervised Patch Label Inference Network with Image Pyramid(WSPLIN-IP)based on the previous work.The method divides the image into patches and designs a weakly supervised learning model to train the patch label inference network using image-level label.In this manner,the loss of pixel information that caused by scaling down the input image as the typical CNN-based method does is circumvented.In addition,WSPLIN-IP introduce image pyramid strategy to fully exploit the scale information during the learning process.(3)We systematically compare the performance of various methods including traditional shallow learning methods,deep learning related methods and the WSPLIN-IP method on the large bituminous pavement disease dataset named CQU-BPDD.The effectiveness and superiority of the WSPLIN-IP method are verified through the analysis of the experimental results.In addition,weakly supervised learning strategy can locate the rough location of the diseased area,which improves the interpretability of the model.Plus,we design and implement a road diseases recognition prototype system based on deep weakly supervised learning and image pyramid.