Research on insulator infrared image segmentation based on GANR-UNet

In order to address the problems of insulator segmentation accuracy caused by the existence of complex background, inconspicuous contrast, and image quality not guaranteed in insulator images taken by aerial photography of transmission line inspection, an optimized segmentation model based on Residual UNet (R-UNet) and Generative Adversarial Networks (GAN) is proposed. First, the R-UNet segmentation network is constructed by pre-training the ResNet18 network on a large dataset and using it as a coding network for the UNet model, which effectively improves the segmentation of insulator strings by utilizing its deep feature extraction capability. Then, combined with the generative adversarial network, the GANR-UNet model is proposed to further improve the insulator segmentation accuracy. Finally, the experimental results show that compared with the traditional R-UNet model, the GANR-UNet model improves the precision (Precision) by $\mathbf{0. 6 3 \%}$, the recall by 1.15%, and also excels in the intersection-to-union ratio (IoU) and other metrics, especially in the extraction of insulator edges. This study provides more accurate technical support for insulator infrared image segmentation, which has important practical application value.