Abstract
In this research, we propose an image montage method to synthesize a natural-looking output image by automatically determining the optimal relative arrangement of two input images of different scenes to montage. In this method, the two input images are relatively arranged so as to overlap each other partially. The overlapping region is searched for the most similar paths in both of vertical and horizontal directions, each of which is the chain of pixels with the most similar color distribution between the two input images. An efficient multiresolution method determines the optimal relative arrangement that has the most similar paths with the highest similarity among all relative arrangements. The most similar path is found by the patch-based image quilting search method, which we have developed based on the pixel-based search method used in the image quilting texture synthesis. This patch-based search method evaluates the similarity between two patches overlapping each other, which come from the two input images respectively. The method finds the most similar path that has the chain of patches with the highest similarity along the path. The similarity between two patches is evaluated using their color histograms. After determining the optimal relative arrangement, the "boundary region" and "copy regions" are determined using the most similar paths of both directions. The two input images are stitched together along the boundary region. The copy region copied from each input image is pasted on the output image as it is. The output image is completed by applying feathering and texture synthesis, or Poisson image editing to the boundary region. A "preserved region" can be assigned in each input image in order to include the region on the output image.
