An accurate registration method on the basis of threshold segmentation and optimized scale-invariant feature transform (SIFT) feature descriptor is proposed for synthetic aperture radar (SAR) images with different bands. Improved 2D cross entropy is adopted for image segmentation, and stable regions are detected with morphological processing. Sobel operator is adopted to compute the gradient magnitudes and orientations of SIFT feature points detected in stable regions. The gradient location orientation hologram (GLOH) is optimized for a more practical use in extraction of a new feature descriptor. The distance ratio and fast sample consensus (FSC) algorithm are introduced to obtain the best match result. Experimental results show that the proposed method can improve the efficiency of the algorithm while achieving high registration accuracy.