Research On The Method Of Transmission Line Image Segmentation And Detection Based On FCIS Model

Transmission lines in the image contains complex foreground and background information,a single test can't pinpoint the precise location information of each target and separate segmentation can't for each class of target recognition,so,the transmission line image segmentation and detection at the same time,to obtain more accurate location information and the category,on the basis of to condition monitoring and fault diagnosis of the transmission line is of great significance.In this paper,the wire,insulator,hardware and tower in the transmission line image are taken as the research object,and the following work is done: Considering that the label of public data set cannot be applied to the transmission line image,the VOC data set was used for reference to construct the transmission line image segmentation and detection data set,which was supported by the data set to reproduce the FCIS model.Considering the inherent characteristics of insulators,conductors and fine fittings within the visual range of transmission lines,the RPN network in the FCIS model is finetuned to change the aspect ratio and size of anchor in the RPN network,so as to improve the positioning inaccuracy of insulators,conductors and fine fittings in the transmission line images.Considering the mismatch between the ROI of the input image in the FCIS model and the corresponding position of the ROI in the feature graph,the closest interpolation method is replaced by the bilinear interpolation method to reduce the loss of feature information caused by the position mismatch.Considering that it is difficult to extract the feature information of fine hardware in the transmission line image,the gradient return algorithm in ROI Align is cited to improve the segmentation and detection accuracy of the model for fine hardware.Based on the constructed transmission line segmentation and detection data set,the FCIS model is reproduced,the Mean Average Precision value of the mAP(Average Precision)was 0.504.After changing the aspect ratio and scale size of anchor in the RPN network,the mAP value was increased to 0.513.After citing the ROI Align algorithm,the mAP value was further increased to 0.5292.