Research On Context-aware Cracked Insulator Detection

With the development of machine learning and computer science,object detection and recognition technology has begun to be used in transmission line visual inspection systems.Nowadays,deep learning has become the mainstream of object detection and recognition.However,the cracked insulator image taken by the UVA not only has a small area,which has few features,but also various changing view angles.These unfavorable factors seriously restrict the deep learning methods such as Faster R-CNN and SSD application.To this end,this thesis proposes the research topic of context-aware cracked insulator detection method,and uses the context background knowledge of the cracked insulator to study the method of cracked insulator image annotation,contextual ROI pooling and tilt correction of the insulator.This thesis mainly studies in the following aspects:Firstly,a method for labeling explosive image based on neighborhood context information is proposed.According to the cracked insulator and its neighborhood are located inside of the insulator,the neighborhood of the cracked insulator has the characteristics of one or two insulator dials,and the blasting piece and the adjacent insulator dial are regarded together as a blasting area.Relabel the cracked insulator training sample,then use the R-FCN method to train and test the sample set.Experiments show that for the cracked insulator image,the detection accuracy of the labeling method is improved by 15.8%.Secondly,a ROI pooling algorithm PR_CoupleNet for neighborhood context information fusion is proposed.Different from the R-FCN detection framework,this thesis uses three ROI branches to extract the features of the cracked insulator.The first branch is the original PS RoI Pooling of R-FCN,and the second branch uses the integral pooling method to calculate the ROI area.The third branch uses the integral pooling method to process twice the area of the ROI neighborhood to take advantage of insulator strings and towers in the neighborhood of the cracked insulator.Then,the pooling result of the latter two branches is channel spliced,and the global feature vector of the target is obtained after convolution,and the local feature vector generated by the first branch are added together to fuse the context information of the cracked insulator.In terms of the small number of cracked insulator data samples,this thesis uses a variety of data enhancement methods to expand the expression of data samples.Experiments show that the detection accuracy of PR_CoupleNet is 10.4% higher than the improved cracked insulator label algorithm.Finally,a rotation-corrected cascade detection network R_Cascaded R-FCN is proposed.First,the insulator string and the cracked insulator detection network are designed separately for the convolutional feature map output by the convolutional network.Then,according to the inclusion relationship between the cracked insulator and the insulator strings,a cascade detection network associated with the context area is constructed.In order to overcome the problem of the posture change of the insulator string,an angle regression full-convolution network Adapt-Res101 which can predict the angle between the insulator string and the horizontal direction,is added to the insulator detection loop and the angle parameter is sent to the convolution map correction module.A convolution feature map of the horizontal correction of the insulator string region is obtained,and then perform the cracked insulator detection.During the training process of the cracked insulator,the GT value of the corrected cracked insulator is updated by using the equal area of the cracked insulator before and after the correction.The network finally achieves end-to-end testing.Experiments show that the detection accuracy of the proposed method reaches 92.5%.Compared with R-FCN,the proposed method has better robustness and generalization.