An Unsupervised and End-to-End Registration Method Using Offset Field and Pseudodata for Video SAR Images

Abstract

Video synthetic aperture radar (video SAR) has been successfully applied in many fields and the registration of the video SAR images has been proven to be a crucial step in their preprocessing. However, video SAR images exhibit more severe image differences because of the unique imaging mechanism and the immature imaging methods. This results in existing registration methods failing to achieve satisfactory registration outcomes for video SAR images. The convolutional neural network (CNN) can contribute to improving registration performance. Nevertheless, CNN-based registration methods must be driven by a large amount of labeled data, which is impractical for video SAR images. Therefore, to tackle these problems, we propose an unsupervised end-to-end deep registration method for video SAR images. First, an end-to-end deep registration model (DRM) is proposed to improve the registration performance for video SAR images. In the proposed DRM, the offset field is utilized to indirectly calculate the registered parameters and we construct a CNN, MUnet, to regress the offset field accurately. We also develop a differentiable H-transform and a differentiable spatial transformation to implement the mapping from end to end while allowing DRM to backpropagate the losses during the training phase. Meanwhile, we borrow intensity-based methods to further optimize the registration results. Furthermore, we propose an unsupervised deep training strategy that can use the generated pseudodata with pseudolabel to train the proposed DRM in the absence of large amounts of labeled data. Experiment results on multiple data demonstrate the effectiveness of the proposed registration method.

Keywords

• Convolutional neural network (CNN)
• image registration
• unsupervised learning
• video synthetic aperture radar (video SAR)