A Study About Detecting Overhead Contact System Insulator On Complex Background Based On Deep Learning

With the development of China,the process of railway modernization is accelerating.Overhead contact system(ocs)insulators play a part in the railway construction.The railway network is across east and west,north and south as the area of our country is quite large.Thus,the environment of the surrounding of the railway is very complex,which includes forest,city,dessert,seaside and so on.The explosion of insulators beyond hostile environment makes them easy to be polluted and broken,which can cause flashover.Once flashover happened,the safety of life and wealth are in danger.So,the real-time detection for ocs insulators is very important.In this thesis,traditional method is applied to detect ocs insulators.There are some steps to detect something in traditional method.First image processing,second image segmentation,third the image detection.In this thesis,a parameter-free image segmentation algorithm is proposed.With the result by the proposed image segmentation algorithm,SURF is used to get the final detection result.The ocs insulators are detected successfully,however the detection results with whose background is complex are poor.To deal with defection of traditional method in detecting ocs insulators on complex background,the application toward ocs insulators using deep learning is studied in this thesis.Focusing on the big data needed in deep learning,origin data are collected as origin database,and origin database is enlarged by affine transformation.Finally database is constructed satisfied with the need in deep learning.Three detection networks are studied in this thesis.The first is Faster R-CNN.First,the network is pre-trained in the COCO database,and then the network is trained on the constructed insulator database.According to experiment comparison,the performance of the network proposed in this thesis is better than the unimproved network.Then,the learning rate of YOLO is adapted,which is a segmented learning rate function.With the adapted learning rate,the network avoids the gradient explosion and can converge at a high speed.At last,in order to applying deep learning in the mobile device,the backbone of SSD is replaced by MobileNet,which is a light net architecture.The proposed method shows no significant defects but has weight file with fewer size.In the end,an ocs insulator detection system is implemented based on browser,which is separated from backend and frontend.By the system,the device can detect ocs insulator in realtime without strong computing capability.According to the result,the effectiveness and feasibility of the proposed method is fully approved.