GENERATION OF HIGH-RESOLUTION IMAGES BY SUB-PIXEL IMAGE REGISTRATION

Abstract

Based on an application in a video conference system, we investigate the problem of generation of high-resolution images by registering multiple images to sub-pixel level accuracy. To generate a high-resolution image of a delicate scene such as an A4 size document by a general-purpose video camera, we capture multiple images and process them by two different approaches: super-resolution and mosaicking. For both approaches, image registration of sub-pixel accuracy is the pre-requisite. The local structure registration method is adopted to implement this task. In the conventional local structure method, only translation and rotation are considered. This transformation model is not enough for our application. The distance between the camera and the hand document is not fixed and this will result in a scale-change between the captured images. Therefore. We develop a new local structure registration method with scale change consideration. The method performs well in its applications in both super-resolution and mosaicking approaches. With an additional variable of scale change, the transformation model is closer to its actual situation. The images are then registered more accurately and the quality of the resulted image is better.