Positional Relationship Unknown Panorama Stitching Algorithm

Although it is easy to get regular digital images as the advance of image acquiring systems, it is hard to get panoramas. As panoramic image mosaic technology can generate panorama using a set of regular images easily, this technology has become increasingly popular in digital image processing.Currently, many automatic image mosaicing systems need to know the arrangement of images. A fully automatic image mosaicing system is proposed in this paper, which does not require the arrangement of images as a prior. This is a feature-based method which using Harris interesting points of images. Based on the user's input images, the method will find the arrangement information and output panorama automatically.The process to match the Harris interesting points is divided into two phases in the paper. The variance normalized correlation is used to match the interesting points roughly in the first phase. The rough match is accurate in some extent because unicity and symmetry are imposed on the correlation. The performance of rough match on different images, and the influence of cylindrical and spherical projection on rough match are given.Genetic algorithm is used to find arrangement information after first phase matching. Rough match result of one arrangement is used as its fitness value in the paper. The probability to obtain the correct answer of genetic algorithm is big through the use of optimized crossover and mutation operators.After the arrangement information has been obtained, the RANSAC method is used to get rid of false matches in the second phase matching. The information of transform of image can be further improved by using the least square method on the pure data.At last, a hat-like weight function is employed to hide the edges of each component images. The final panorama is generated successively by adding new images to the existing panorama one at a time.