A lightweight YOLOv5 model for small insulator defect detection

作者：Na Liu, Fangzheng Peng, Sheng Hua, Chonghao Yue, Ligang Dong · 发表于：Other Conferences · 年份：2024 · DOI：10.1117/12.3024278 · 被引用次数：2 · 研究领域：Engineering

To address the issues of low accuracy and poor real-time performance in insulator defect detection in power transmission systems, we propose a lightweight insulator defect detection model based on YOLOv5. Firstly, we have introduced a separable vision transformer (SepViT) module to replace several CBS and C3 modules in the final layers of the backbone network, which can greatly reduce network computational complexity and improve detection speed without losing detection accuracy. Secondly, we introduced a bidirectional feature pyramid network (BiFPN) to address the issue of information loss in the original network structure. Finally, we replaced the CBS module in the Neck network with an improved DSC module to reduce the network load again and introduced an Involution block to compensate for the problem of DSC channel information loss. The experimental results show that our model has achieved better performance compared to other insulator detection models, especially for the original YOLOv5s model. Our detection accuracy has been improved by 6.6%, the model size has been reduced by 24.3%, and the detection speed has been improved by 240%, reaching 70.4fps.