Abstract：The detection of composite insulator defects in substations still relies on inspection by operators，

which is a heavy workload and prone to leakage due to visual fatigue. To reduce the computational resource

consumption and shorten the training time，the region convolutional neural networks（RCNN）was improved by

reorganizing the convolution kernel，and a detection method was proposed for insulator crack shape features. The

method can meet the premise of insufficient training sample data，but also can get better convolutional neural

networks（CNN）training effect，and finally achieve accurate crack recognition. In the training phase，the RGB

three-channel decomposition method was used to expand the training data set，the median filtering method was used to remove the noise，the improved convolutional kernel was used to train the CNN. In the test phase，the images were decomposed by RGB three-channel decomposition and input to CNN to get the exact crack center coordinates and length. The non-maximum suppression（NMS）algorithm was used to de-weight the images to get the final crack recognition results. The example analysis shows that the propose method can still achieve good recognition accuracy，and accurately identify the specific location of cracks under the premise of insufficient training samples.

      Key words：insulator defect identification；convolutional kernel；image processing；region convolutional neural networks（RCNN）；RGB three-channel filtering