Abstract: [Purposes]To address the challenges of noise-susceptible backgrounds and small,irregular discharge regions in ultraviolet (UV) imagery—factors that undermine the accuracy and robustness of traditional approaches—and to improve both detection/segmentation performance and engineering applicability for corona discharge.[Methods]We augment the U-Net framework with atrous spatial pyramid pooling to enhance multi-scale feature representation and embed a convolutional attention module into skip connections to highlight discharge regions while suppressing background clutter.A composite loss combining binary cross-entropy and Dice loss balances pixel-level accuracy and region overlap.The model is trained and validated on a UAV-acquired UV dataset of 220 kV substation equipment.[Findings]Experimental results show that the proposed method achieves an mIoU of approximately 0.83 and a Dice score of approximately 0.89 on real data, with precision, recall, and inference speed of 0.875, 0.935, and 1.99 frames/s, respectively, demonstrating a favorable balance among segmentation accuracy, target coverage, and real-time performance.[Conclusions]The improved U-Net strikes a favorable balance between accuracy and efficiency,offering a reliable technical solution for automated and intelligent inspection of corona discharge in power equipment.

Key words: corona discharge detection, ultraviolet image segmentation, smart grid inspection, multi-scale feature fusion, attention mechanism