Abstract:

UAV images for powerline inspection usually have complex backgrounds, and often contain a lot of small targets, which may lead to a high rate of missed detections and false detections when processed by the general feature extraction networks for object detection. To address this, a feature focus & diffusion network (FFDN) was proposed for feature fusion, and an improved algorithm (YOLOv8-SFD) based on FFDN and YOLOv8 was designed for powerline component detection. Spatial-to-depth non-stride convolutions (SPDConv) were employed in the backbone network instead to preserve small-scale features and reduce feature loss caused by stride convolutions. The traditional feature pyramid network was replaced with the proposed FFDN. At the feature fusion stage, the feature focus modules in the FFDN were utilized to expand the receptive field and fuse multi-scale features, and the output feature maps by them were then diffused across different scales to enhance small target detection accuracy. Finally, the original YOLOv8 head was replaced with a dynamic detection head (DyHead) that integrates three attention mechanisms (scale, space, and task), to further enhance the performance. Experimental results demonstrated that YOLOv8-SFD achieved an accuracy rate of 76.7%, which was 7.6% higher than YOLOv8n; a recall rate of 43.0%, which was 2.0% higher; and a MAP of 48.2%, which was 3.8%. YOLOv8-SFD effectively enhanced the detection precision for small and obscured targets, and the detection speed reached 119 FPS, which satisfied real-time detection requirements.

Key words: powerline inspection, feature fusion, object detection, feature focus & diffusion network, YOLOv8