For overcoming the low image registration of the SIFT(scale invariant feature transform) when it is applied to SAR image registration, put forward the SAR image accurate registration algorithm. At first, the method extracts the SIFT descriptor and improved rotation invariant textured feature descriptor, then use canonical correlation analysis feature fusion method to fuse 2 descriptors and create new feature descriptor. Next the improved weighting distance among every feature point in 2 images is calculated and introduced into affine transform model for transform parameter and realize image registering. The simulation results show that registration result of the proposed method under much better than SIFT method and PCA-SIFT, and registration precision reaches sub-pixel level when image in conditions of scale variation, grey scale variation, and rotation angle variation.