DFW-YOLO: A small insulator target defect detection algorithm based on improved YOLOv8s

With the continuous progress of deep learning technology, UAV aerial photography faces significant challenges for insulator defect detection. Aiming at the problems of low detection accuracy of existing target detection algorithms and difficulty in recognizing small target defects, we propose an improved small target insulator defect detection algorithm based on YOLOv8s, named DFW-YOLO. Firstly, the Detect_Efficient lightweight detection header is proposed using partial convolution (PConv) to lighten the original detection header. Secondly, a FocalModulation focal modulation module is introduced into the backbone network to enhance the model’s extraction and fusion capabilities for features at different scales. Finally, to enhance the model’s focus on poor-quality samples and reduce the harmful gradients they produce, a loss function with a Wise-IoU V3 dynamic non-monotonic focusing mechanism is used instead of the original CIOU loss function. We conducted experiments on a publicly available dataset of UAV aerial photography. According to the experimental data, DFW-YOLO achieves an 86.8% mAP in insulator defect detection, showing a 6.8% improvement compared to the original YOLOv8s model and generally exceeding the performance of other prominent models. Utilizing this method can effectively boost the accuracy of identifying insulator defects in small targets.