Hydrophobicity Classification of Composite Insulators Based on Image Enhancement and Deep Learning

Accurately judging the hydrophobicity classes (HCs) of the composite insulator helps to grasp the antipollution flashover performance of the insulator and prevent accidents in time. Aiming at the problem of low HC recognition accuracy of hydrophobic images and interference of operators’ subjectivity in the feature extraction process, an intelligent HC recognition model of composite insulators based on the Contrast Limited Adaptive Histogram Equalization (CLAHE) image enhancement method and ResNet deep learning network is proposed in this paper. Firstly, composite insulators under varying HCs were simulated by spraying ethanol solutions with different volume fractions, and a series of hydrophobic images under experimental conditions was obtained. To improve the generalization ability of the model on photographic factors (photographing distance, light intensity, insulator color), images were preprocessed with gray processing, cropping, enhancement, and so on. The transfer learning strategy was used for training and verifying the pre-trained ResNet network by applying the preprocessed image dataset so that it is suitable for recognizing the HC of composite insulators. The results show that the CLAHE-ResNet18 model proposed in this paper can still effectively identify each HC under varying photographic factors and complex conditions. This model can maintain an accuracy above 96.74% in the testing dataset, with good generalization ability and application value.