Research On The Localization Methods Of Insulator In Infrared Images Based On Deep Feature Representation

This paper focuses on the problems that most of the electrical insulation equipment inspection systems can not locate the object automatically.Firstly,this paper presents an infrared on-line inspection and warning system to detect the thermal fault of transformer bushings in the early stage.Then,several feature representation generation methods are proposed based on the insulator localization task.The novel contributions are summarized as follows.1.For real applications,an on-line inspection and warning system based on infrared image is designed and applied,and an infrared insulator image database is built up from the acquired images.2.This paper mainly studies the insulator feature representation problems.Based on Binary Robust Invariant Scalable Keypoints(BRISK),a mid-level feature representation for infrared insulator string is proposed.This paper has utilized the binary local features to make a more semantic representation.This paper also takes a large step from feature engineering to feature learning by introducing the deep learning theory into insulator recognition task.The method to aggregate deep convolutional feature maps for insulator detection in infrared images is proposed.This paper also delves deep into the convolutional feature maps for robust representation for insulator images.In order to squeeze the remaining representation power of deep convolutional neural network,this paper delves deep into the convolution layers,and dig the multi-scale hierarchies for compact image feature aggregation.The activation pattern of deep neurons are studied.3.An insulator localization method based on BRISK feature matching is proposed,and this automatic recognition method overcomes the negative effects of deviation caused by mechanical transmission;Then,in order to get rid of the templates and acquire higher localization accuracy and generalization ability,an insulator localization method based on multi-scale mid-level representation is proposed.This paper takes advantages of the deep feature maps for robust representation,and applies object proposal for insulator localization.4.Based on the powerful generalization and representation ability of deep features,a status classification method of an insulator is proposed.This paper proposes a multi-patch deep feature generating methods to representation normal,defect or cracked insulators.The proposed multi-patch deep feature extraction method improves the classification accuracy from 91.3737% to 98.0687%,which has outperformed the hand-crafted features by a large step.