Resolution enhancement of image sequence using motion-compensated sub-pixel registration and interpolation

In many applications of image processing, it is necessary to reconstruct a single, high resolution image from multiple low resolution images of the same scene. To make a high resolution image from multiple low resolution images, relative shifts between images need to be estimated accurately enough so that different information from images can be extracted and combined. Obtained by a moving image capturing device, low resolution images are temporally-correlated by spatial shifts provided that local object motions are compensated. For most image processing algorithms which deal with image sequences, motion compensation is an essential step prior to the main algorithm. Different methodologies for motion estimation are reviewed and their limitations are examined.; To estimate relative shifts between images, a sub-pixel registration algorithm using differential method with edge line model is proposed. For the pixel unit registration, it uses edge chain code to find a best match between two images. After the pixel unit shifts are compensated, parallel processing to estimate each sub-pixel shifts is performed. Previous techniques on sub-pixel registration are reviewed. Experimental results on simulated images are compared with other algorithms.; With estimated sub-pixel shifts, a higher resolution image is reconstructed by the interpolation algorithm. Previous techniques with single input and multiple inputs are reviewed. The algorithm is free of ill-posedness since the number of inputs are not constrained. The initial guess of a high resolution image is taken to be the weighted sum of two dimensional reconstruction filter output and it is refined in an iterative fashion by minimizing the error between the input low resolution images and images from simulated imaging process in which directional gradient information is used for the selection of low resolution images. For the case of doubling the image size for both directions, interpolated images using the algorithm are made and compared with images from single input interpolation algorithms.