Pavement Disease Detection Based On Transfer Learning

The timely and effective detection of road pavement diseases is essential to ensure the safety of road transportation.At present,manual detection and semi-automatic detection equipment are the main detection methods of pavement diseases,which are high costly and inefficient.Therefore,the automatic detection method of pavement disease based on digital image has become a popular research topic.In this thesis,we classify pavement diseases and locate the diseased area by computer vision related technologies such as image adaptive enhancement,weight adaptive adjustment,transfer learning and domain adaptation.The detail is as the following two aspects.(1)Research on multi-classification of pavement diseases based on image adaptive enhancement,weight adaptive adjustment and parameter migration.For addressing the uneven distribution of gray level,we first enhance the patch split from the whole image,then enhance the image composed of enhanced patches.this kind of enhancement takes local enhancement and global enhancement into consideration.For the problem of unbalanced data distribution,an adaptive weight adjustment method is proposed to improve the detection ability of pavement disease by adaptively modifying the weight of disease categories.Besides,we adopt the transfer learning method of parameter transfer to reduce the training cost and improve the generalization ability of model.Experiments show that the F1 score of the disease category can increase up to 22% and the classification accuracy can increase 2.02% by adaptive enhancement and adaptive weight adjustment.(2)Research on weakly supervised pavement disease area detection based on domain adaptation.CQU-BPDD is a multi-classification dataset of pavement diseases,which lack of location information of disease area.So,we cannot detect the disease area after getting the detail disease category of pavement.In order to detect the pavement disease area on the condition of weakly supervision,the knowledge in the RDD2020 domain having pavement disease area annotations is transferred to the CQU-BPDD domain through domain adaptive methods.Specifically,an intermediate domain including the feature space,category space and conditional probability distribution space of source domain and target domain is constructed.Then we migrate the YOLOv5 model on the intermediate domain to realize domain adaptation.Experiments show that domain adaptation improves the accuracy by 11%,and the detection F1 score of the disease category is in-creased by up to 21% compared with direct domain migration.In addition,the labeling information provided by domain adaptation also can be considered as reliable data support for future pavement disease area detection research.The research on solving pavement disease detection under weak supervision through domain adaptation provides a research idea for data automatic labeling problems.