SPB-YOLO: An Efficient Real-Time Detector For Unmanned Aerial Vehicle Images

Recently, using unmanned Aerial Vehicle(UAV) to capture images has become a popular application. However, the large scale variation and dense object distribution characteristic of UAV images brings challenges to object detection. Hence, we propose an efficient end-to-end detector named SPB-YOLO for UAV images. In this paper, firstly we design a Strip Bottleneck (SPB) module to better understand the width-height dependency by using an attention mechanism for improving the detection sensitivity of different scales' objects in the UAV image. Secondly, we propose an upsample strategy based on Path Aggregation Network(PANet) for the feature map and add another one detection head compared to YOLOv5, which specially deal with the detection task of dense objects distribution. Finally, we execute some experiments on two public datasets, and the results show that the proposed SPB-YOLO outperforms other latest UAV image detectors and makes a good trade-off between detection accuracy and speed.