For the accurate registration of high resolution Synthetic Aperture Radar (SAR) images, local transformation functions are preferable to global ones. In this paper, a local mapping model, namely neighborhood reconstruction model, was presented to fit the geometric distortions in SAR image registration. First, each point in the reference image was characterized by the neighboring Control Points (CP) with reconstruction weights. Then, a criterion for neighboring CP selection was proposed to minimize registration error at each point. Finally, the associated point in the sensed image was resampled according to the geometric distortions and the reconstruction weights. The theoretical support from neighborhood reconstruction model to the classical piecewise linear approach was also presented. Experiments on both simulation data and real high resolution SAR images show that the registration accuracy is improved.