Fast Registration of High-Resolution Optical-SAR Images Based on Grayscale Normalized Mutual Information

Due to their different imaging principles, optical and SAR images can reflect complementary information, and the fusion application of the two can improve the observation ability of ground remote sensing. The registration of heterogeneous images is a prerequisite for the fusing of different remote sensing images. Optical and SAR images are highly heterogeneous. For high-resolution images, the local similarity between SAR and optics is even worse, with extremely limited available information. Therefore, how to find similarity information when the image contains less information is a challenge in the field of image registration nowadays. The existing registration algorithms face the problems of poor registration accuracy and long registration time when registering high-resolution heterogeneous images. Therefore, this paper proposes to extract the correlation between high-resolution heterogeneous images from a small amount of information, and perform accurate and fast automatic registration of small-scale images near the target of interest. Considering the local distortion of SAR and optical images, which results in poor consistency of large images, independent registration within a local range is relatively suitable. Therefore, based on the use of grayscale normalized mutual information as a similarity measure, this article proposes an innovative segmentation registration approach that achieves overall registration through each block registration. And the strategy of combining large and small blocks effectively solves the contradiction between the low amount of information in small images and the high overlap of information in large images. The experimental results show that the overall registration success rate of optical-SAR images is significantly improved by using the registration strategy of combining larger blocks and smaller blocks based on gray normalized mutual information. Moreover, as it does not require traditional methods such as feature extraction and feature matching, it can achieve fast automatic registration.