Abstract: Aiming at the characteristics of infrared images with low contrast, complex backgrounds, and blurred contour information, this paper proposes an improved infrared image segmentation model, UNet++, for power equipment based on UNet. The method makes full use of the encoding-decoding structure of UNet, on this basis, redesigns the dense connections and jump paths, and deepens the hierarchy of the “U” network. The network model is pruned so as to extract more detailed features from the image while avoiding the decrease of inference speed. The experiments are trained and validated using UNet++, and the results show that the average segmentation accuracy can reach more than 88%. The study shows that applying image segmentation technology to infrared scenes can reduce the tedious process consumed by traditional manual judgment devices, achieve fast and highly accurate analysis of massive data, and finally reach the goal of automatic segmentation.