Abstract: Segmenting discharge spots in ultraviolet images of electrical equipment helps to quickly locate faulty areas and assess the discharge intensity, thereby providing technical support for maintaining the safe operation of power grid systems. The irregular shape and fuzzy edges of the ultraviolet discharge spot can easily lead to missegmentation and missing segmentation. A DTDU-Net UV discharge spot segmentation method based on an improved U-Net is proposed. First, a residual structure and deformable convolution were introduced into the encoder to enhance the feature extraction capability and reduce missing segmentation. Second, the U-Net skip connection was replaced with a channel cross-fusion transformer to effectively capture the cross-channel interactions and improve spot mis-segmentation. Finally, in the decoder part, an ultra-lightweight dynamic up-sampler, DySample, is used to replace the original up-sampling operation, which can better retain the image details and alleviate the problem of missing segmentation. The experimental results showed that the average crossover ratio of the improved network for UV spot segmentation was 95.17%, and the average accuracy was 96.79%, which were improved by 6.32% and 6.77%, respectively, compared with those of U-Net. Additionally, the segmentation effect was good.