Aircraft swarm detection based on multi-scale-feature-fused attention

Abstract

To improve the detection performance of aircraft swarms in remote sensing images with characteristics of small target, scale diversity and dense distribution, a multi-scale-feature-fused attention mechanism is proposed and used for deep learning networks in this paper. Based on the fundamental YOLOv5 network, an enhanced multi-scale CBAM attention module that combines the channel attention and the spatial attention is performed on the fused feature maps at various stages and scales. Consequently, more detailed attention information can be obtained. Experimental results demonstrate that the proposed method can effectively improve the detection accuracy of aircraft swarm targets compared with some traditional methods. In detail, the proposed method can reach 9.2% higher recall than the original YOLOv5 and 3.4% higher recall than the YOLOv5 integrated with traditional CBAM modules while the computational cost is similar