Multi-scale Object Detection in Remote Sensing Images Based on Dual Attention Mechanism

Abstract

Aiming at the detection difficulties in optical remote sensing images, such as the large range of target scale variation and the diversification of target directions, a remote sensing image target detection algorithm based on dual attention mechanism and multi-scale feature fusion is proposed. On the basis of the lightweight network Mobilenet v3, a dual attention mechanism is added, which brings obvious performance gains to the detection model while adding a few parameters. Through multi-scale feature fusion, the receptive field is enlarged, the fusion features of targets of different scales are obtained, and the multi-scale target detection effect is effectively improved. Anchor box detection in any direction is performed for multi-directional targets by introducing a rotating frame. Experiments show that the proposed method has higher detection accuracy and faster detection speed than the current popular methods.