EdgeAware YOLO: Enhancing Small Object Edge Detection in Base Station Equipment Inspection

The small object detection task has always presented a significant challenge in the field of object detection research. However, existing methods often struggle to effectively focus on the blurred edges of small objects, leading to overlapping predicted bounding boxes or inaccurate predictions for similar small objects in dense scenarios. In this paper, we propose a novel small object detection network based on the YOLO framework, called EdgeAwareYOLO, to alleviate this challenge. In order to focus on the edge information of small objects and improve the location ability of dense small objects, we propose a novel module called channel attention and spatial attention fusion module. Additionally, we propose Side and Corner Aligned Intersection over Union Loss function to alleviate the issue of gradient vanishing caused by low overlap between predicted bounding boxes and ground truth boxes based on traditional Intersection-over-Union loss and facilitate bounding boxes regression in cases of high overlap. Extensive experiments demonstrate that the proposed EdgeAwareYOLO achieves significant improvements in small object detection by capturing edge details and generating more reliable results. The implications of the proposed method hold promise for enhancing base station equipment inspection.