In the process of reconstructing super-resolution images by POCS algorithm, Image registration and initial estimation have important influence on the result of reconstruction.For solving the problem that the small block matching is liable to be interfered by noise and the big block matching can’t track the local moving objects accurately, a method to determine the spatial flatness of the image and divide the blocks according to the flatness of the image is provided.It can change the block size dynamically;As the initial estimation for the single frame is lack of enough information, we use all known low-resolution image to register the interpolated image.Then the initial estimation is generated by the optimal interpolation method.Experiments show that the improved matching algorithm reduces the influence of noise and local motion on block-matching accuracy and it’s also very sample.The new algorithm also improve the PNSR of the initial estimation and speeds up the convergence.