Unsupervised Image Registration for Video SAR

Existing approaches for SAR image registration focus on the global transformation correction between images. However, there are often local deformations Due to time-changing viewpoint of video SAR, images suffer a lot from deformations, which can result in false alarms moving target detection. This article presents an unsupervised approach use detection, has good performance and acceptable processing efficiency. The designed learning-based framework is cascade two convolutional neural networks. first network directly predicts parameters rigid reference unregistered images, recovers them. Then, second uses registered as input then displacement field. After that, we put limitation predicted field prevent shadows being aligned. Finally, with used compensate Processing results real have shown proposed convincing generation ability.