Dual efficient self-attention network for multi-target detection in aerial imagery

Aerial imagery target detection has been widely used in the military and economic fields. However, it still faces a variety of challenges. In this paper, we proposed several efficiency improvements based on YOLO v3 framework for getting a better small target detection precision. Firstly, a dual self-attention (DAN) block is embedded in Darknet-53's ResNet units to refine the feature map adaptively. Furthermore, the deep semantic features are cascaded with the shallow outline features in a feedforward deconvolutional module to obtain context details of small targets. Finally, introducing online hard examples mining and combining Focal Loss to enhance the discriminating ability between classes. The experimental results on the VEDAI aerial dataset show that the proposed algorithm is significantly improved in accuracy compared to the original network and achieves better performance than two-stage algorithms.