A Quasi-feature based Image Mosaic Algorithm Using Modified Block Matching Criteria

Abstract

It is important to know the projective transformation between two images in order to construct the image mosaics with multiple images. The conventional methods to obtain the projective transform use the geometrical features or the intensities of the image. These methods execute the iterative computation, so that the total amount of computation to extract the correspondence and to induce the parameters of the transform is too much.
In this paper, we propose the projective registration algorithm using the quasi-feature points, which are based on the intensity. We define the quasi-feature point as a central pixel of a block that has enough texture information. To determine these feature points, we extract the overlap image between two images using the Fourier method and divide this overlap image of the reference into the four subareas. Then we choose the central pixel of the block which has the maximum local variance in each subarea. We use the block matching algorithm, which is considered of both the camera movement and the variation of the illumination conditions, in order to match these four feature points with the four points in the target image. The proposed algorithm is applied to various images to estimate the performance and the simulation results present that this method finds the correct correspondence.