Aiming at the problems of low performance matching, more mismatching pairwise, and low registration precision, which are the characteristic of traditional SAR image registration methods, we propose a novel and efficient local invariant feature-based algorithm. First, the feature points are detected by features from accelerated segment test(FAST) method and described by DAISY descriptor in SAR image. Second, Kd-tree-based dual-matching strategy and random sample consensus (RANSAC) are used to establish fine feature matching. Third, affine transform model is estimated for image resampling and transformation, and rough registration is implemented. Finally, feedback mechanism is constituted for fine registration based on the estimation of registration precision. The flexibility and efficiency is demonstrated by experiments with slant range SAR images acquired from different working model, different times, viewpoints, wavelengths and polarizations.