Polarimetric SAR Image Affine Registration Based on Neighborhood Consensus

Abstract

As the base of Synthetic Aperture Radar (SAR) image processing, the registration of polarimetric SAR images requires high accuracy and a fast speed. Most methods used to register polarimetric SAR images based on deep learning are combined with patch matching and iterative estimation, e.g. the random sample consensus algorithm. However, end-to-end deep convolutional neural networks have not been used in the non-iterative affine registration of polarimetric SAR images. This paper proposes a framework for end-to-end polarimetric SAR image registration that is based on weakly-supervised learning and uses no image patch processing or iterative parameter estimation. First, feature extraction is performed on input image pairs to obtain dense feature maps with the most relevant k matches kept for each feature point. To filter the matched feature pairs, the 4D sparse feature matching maps are then fed into a 4D sparse convolutional network based on neighborhood consensus. Lastly, the affine parameters are solved by the weighted least square method according to the degree of confidence of the matches, which enables the affine registration of the input image pair. As test image pairs, we use farmland data from Wallerfing, Germany obtained by the RADARSAT-2 satellite and Zhoushan port data from China obtained by the PAZ satellite. Comprehensive experiments were conducted on polarimetric SAR image pairs using different orbit directions, imaging modes, polarization types and resolutions. Compared with four existing methods, the proposed method was found to have high accuracy and a fast speed.

Keywords

• neighborhood consensus
• affine transformation
• polarimetric sar
• image registration
• sparse convolutional neural network