Composite Insulator Defect Identification Method Using Improved RCNN Convolution Kernel

The detection of composite insulator defects in substations still relies on manual inspection. In this paper, we propose a method for crack shape features by improving the RCNN convolution kernel. can meet premise insufficient training sample data, but also get better CNN effect, and finally achieve accurate recognition. phase, RGB three-channel decomposition is used to expand data set; median filtering remove noise; improved convolutional kernel train CNN; test images are decomposed input exact center coordinates length; NMS algorithm de-weight final recognition results. example analysis shows that paper good accuracy accurately identify specific location cracks under samples.