Moving Target Shadow Detection In Video SAR Based on Multi-Frame Images and Deep Learning

Moving target shadow (MTS) can reflect the real position and state information of moving targets. However, the change of scene and motion state of moving targets will lead to small Contrast Ration of a Distributed Target (DTCR) and large Extension of Shadow (SE) in Video SAR images, which leads to increasing the difficulty of MTS detection. To solve this problem, the multi-frame detection method, memory enhanced global-local aggregation (MEGA), is applied to Video SAR field for the first time and improved (I-MEGA) in our work. Firstly, global semantic information of each frame is extracted by feature pyramid network (FPN) and proposals of adjacent frames are generated by RPN to provide local localization information. Then, the global and local information are aggregated to key frame through the relation module, so as to enhance the feature of the MTS in key frame; Finally, the aggregated feature map is fed to the detection network to obtain the final detection result. Compared with other single-frame methods, I-MEGA greatly reduces the false alarm rate and effectively improves the ability to detect MTS in Video SAR with small DTCR and large SE on the dataset formed of the real data published by Sandia National Lab (SNL).