This paper proposes a semantic segmentation model called VA-Unet, designed to address the challenges of complex backgrounds, slight spot separation, complex feature selection, and low segmentation accuracy encountered in ultraviolet (UV) detection tasks of electrical equipment. VA-Unet incorporates the VGG16 feature extraction module and transfer learning to accelerate training and enhance the model's generalization capability. Additionally, an Attention Gate is integrated to improve segmentation precision by focusing on relevant features, enabling accurate detection of UV discharge spots in images. To address the issue of sample imbalance in the UV discharge spot dataset, VA-Unet employs a hybrid loss function in place of a conventional single loss function. Experimental results demonstrate that VA-Unet achieves superior performance in the precise localization and accurate segmentation of UV discharge spots. The model attains an IoU of 84.09%, PA of 88.20%, and F1-score of 91.35%, representing improvements of 14.41%, 3.24%, and 9.22%, respectively, compared to the baseline U-Net model.