Dim and small objects detection in aerial images with stacked attention mechanism and improved loss function

Though dim and small objects detection in aerial images has been developed recently, there are still some challenges that have not been addressed such as background inundation, dense distribution, various scales, and multi-direction. In addition, many existing detection algorithms bear heavy computing load. To alleviate the above problems further, we improve the YOLOv5s in the following aspects: first, GIoU regression loss is replaced with weighted loss combining IoU loss and NWD loss to reduce the influence resulted from the objects with various sizes and densely distributed; second, cascaded and differentiated attention mechanisms are introduced into the different modules of YOLOv5s to distinguish the objects from the complex background; third, common convolution is modified with Ghost convolution to reduce the number of parameters and GFLOPs, building a more lightweight model, to enable the deployment on the embedded devices with limited hardware resources; and finally, we experiment on DIOR and DOTAv1.0 to investigate the effectiveness and efficiency of the improved YOLOv5s, experimental results show that the improved YOLOv5s achieves the satisfactory trade-off between detection rate and speed compared to the other state-of-the-art works. DIOR and DOTAv1.0 can be downloaded at: http://www.escience.cn/people/gongcheng/DIOR.html,https://captain-whu.github.io/DOTA/dataset.html.