Abstract: Surface defects of insulators can severely deteriorate their performance and may even lead to power failures. Therefore, accurate identification and precise localization of surface defects on insulators are essential for ensuring the safe operation of the power system. The focus of insulator defect detection is the accurate identification and localization of defects in its materials. However, conventional deep learning-based methods face limitations such as poor detection rates when identifying small defects. To address this issue, this paper proposes aglobal and local dual-branch attention module (GLDA). Firstly, the global branch utilizes a traditional attention mechanism but applies downsampling to K and V to reduce computational load, capturing low-frequency global information. The local branch effectively integrates shared weights and context-aware weights to aggregate high-frequency local information. Secondly, an Adaptive Spatial Feature Fusion (ASFF) module is used to extract multi-scale information of defects, enhancing detection accuracy. Lastly, the CIoU algorithm is introduced to optimize the training loss of the baseline model. Experimental results show that the proposed method achieves accuracy, recall, and mAP of 93.2%, 92.9%, and 95.2%, respectively.

Key words: insulator defect detection, deep learning, attention mechanism, multi-scale