An Improved Deep Learning Network for Spray Detection

Hydrophobicity is an important indicator to characterize the aging status of composite insulators. Conventional computer vision methods face significant challenges in on‐site hydrophobicity detection of composite insulators due to glare interference on the spray. To address this challenge, DelightNet, a novel dual‐stream semantic segmentation network which learns separately from the normal spray images and detail layers of spray images to accurately detect and track drops on insulator and improves semantic segmentation networks to mitigate the influence of glare interference, is proposed in this paper. Several experiments were conducted to evaluate the accuracy of DelightNet, and DelightNet has higher average ‐score and average mIoU than several kinds of classical deep learning network: the proposed DelightNet achieves an average mIoU of 0.725 and average score of 0.809, outperforming state‐of‐the‐art methods by 2.4% and 1.4%, respectively. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.