A Lightweight Object Detector Based on Spatial-Coordinate Self-Attention for UAV Aerial Images

Object detection is one of the most widespread applications for numerous Unmanned Aerial Vehicle (UAV) tasks. Due to shooting angle and flying height UAV, compared with general scenarios, small objects account a large proportion aerial images, common object detectors are not extremely effective in images. Moreover, since computing resources UAV platforms generally limited, deployment number parameters on difficult. This paper proposes lightweight detector YOLO-UAVlite Firstly, spatial attention module coordinate modified combined form novel Spatial-Coordinate Self-Attention (SCSA) module, which integrates spatial, location, channel information enhance representation. On this basis, we construct backbone, named SCSAshufflenet, combines Enhanced ShuffleNet (ES) network proposed SCSA improve feature extraction reduce model size. Secondly, propose an improved pyramid model, namely Slim-BiFPN, where new convolutional blocks loss during map fusion process while reducing weights. Finally, localization function increase bounding box regression rate improving accuracy. Extensive experiments conducted VisDrone-DET2021 dataset indicate that, YOLOv5-N baseline, reduces by 25.8% achieves gains 10.9% mAP0.50. Compared other detectors, both mAP improved.