MAEE-Net: SAR ship target detection network based on multi-input attention and edge feature enhancement

Abstract

Synthetic Aperture Radar (SAR) imagery has a wide range of applications in search and rescue ships lost contact and military reconnaissance. When detecting multi-scale targets, better determination of the target edge is conducive to improving the detection accuracy of the model, but most of the existing methods lack research on this aspect. To fix the problems mentioned earlier, this paper suggests using a SAR Ship target detection network called MAEE-Net. In this paper, a multi-input attention-based feature fusion module (MAFM) and an edge feature enhancement module (EFEM) are proposed. MAFM uses attention mechanism with multi-input and multiple-output to improve attention to shallow feature map target and suppress invalid information, so as to improve the information utilization rate of each layer. To make the network better at detecting the edges of ships, EFEM uses double-branched structure to carry out fine-grained information retention and edge feature extraction. PIoU v2 is introduced to enhance multi-target processing capability. Experiments were carried out on SSDD dataset and SAR-Ship-Dataset, the overall detection accuracy was as high as 98.6% and 94.7%. The detection accuracy was 93.5% and 99.3% on inshore and offshore sub-datasets of SSDD dataset. Experimental results on two datasets show that our model is impactful.