Subpixel shift estimation in noisy image using a wiener filtered local region

Image registration is a fundamental technique in image processing that requires a precise shift estimation with subpixel accuracy. However, when an image is affected by noise, aliasing, or other artifacts, this increases the estimation error. Accordingly, a novel accurate subpixel shift estimation method is proposed for the registration of noisy images. In the proposed method, instead of using the whole noisy image, a local region is selected based on analyzing the gradient and distribution of the frequency component in the vicinity of a corner. Next, an adaptive Wiener filter, where the window size is varied depending on the property of the local region, is applied to suppress remaining noise. Finally, an iterative phase correlation of the selected local region is performed within a restricted range using a coarse to fine method. In experiments, the proposed method provides a higher estimation accuracy than other methods. Plus, when applying the proposed method to the registration of super resolution, an enhanced edge sharpness is seen in the resulting images.