MDS-YOLO Model-Based Defect Detection Method for Porcelain Insulators Using Infrared Images

Abstract

With the rapid development of image processing technology in recent years, the detection of insulator defects through infrared images has become an important online inspection technology. In practice, the insulator infrared image shooting needs to deal with shooting angle, background complexity, and other issues that decrease the detection accuracy. Also, small targets are difficult to identify and the detection of defects remains a problem. In order to solve these issues, this paper proposes a small target multiple defects YOLO algorithm. Based on YOLOv8, a hybrid model of self-attention and convolution is used to aggregate convolution and self-attention. Then efficient convolutional network (EfficientNetV2), is applied to improve the training speed of the model and the parameter efficiency, to ensure that the model is lightweight as a whole. And adopting a bi-directional feature pyramid network to improve accuracy through multi-level feature pyramids and bi-directional information transfer. The multilevel feature pyramid and bidirectional information transfer are adopted to improve the precision. Finally, the inner-SIoU loss function is used to improve the recall and precision of the small targets and enhance the robustness of the model to small targets. In order to obtain test data, this paper conducts defective insulator infrared image experiments to obtain infrared images under different conditions. After experimental verification, the MDS-YOLO algorithm proposed in this paper achieves an average of 87.85% mAP and 6.0 GFLOPs, which meets the requirements of recognising defective insulators with small targets and the effectiveness and superiority of the algorithm proposed in this paper are proved by ablation and comparison tests.