Abstract:

This paper proposes an automated registration method for multispectral remote sensing images to address the significant non-linear intensity differences among such images. The proposed method first introduces the local self-similarity (LSS) feature, and builds the feature descriptors capturing the internal geometric layouts and shape properties of images in the log-polar coordinate. Then, a shape similarity measure is defined by combining LSS and normalized correlation coefficient, followed by the template-matching scheme to achieve image registration. Experimental results show that the proposed method is robust to non-linear intensity differences among images and achieves reliable registration accuracy，and obtains a higher correct matching rate in consideration of the computational efficiency compared with state-of-art registration methods.

Key words: multispectral remote sensing image, automatic registration, LSS, mutual information, correct matching rate