UV image discharge spot segmentation for electrical equipment based on improved U-net

This paper proposes a semantic segmentation model called VA-Unet, which aims to solve the problems of complex background and slight spot separation, complex feature selection, and low segmentation accuracy faced by traditional segmentation methods in UV detection tasks of electrical equipment. VA-Unet introduces the VGG16 feature extraction module and migration learning to improve the training speed and enhance the model's generalization ability and adds the Attention Gate to improve the model's accuracy to achieve accurate segmentation of UV image discharge spots. In addition, VA-Unet uses a hybrid loss function instead of a single loss function to solve the problem of sample imbalance in the UV discharge spot dataset. Experiments show that VA-Unet model has outstanding performance in positioning and segmentation of ultraviolet image discharge spots, and its IoU, PA and F1-score evaluation indexes have increased by 14.41%, 3.24% and 9.22%, compared with the initial U-net network.