Research And Application On Registration In Image Mosaic Based On SIFT

Image Mosaic is a technology that takes a series of images which are overlapped with each other and taken from the same scene into a seamless and high quality image which has a wide view. In contrast with specialized equipment, Image Mosaic can get the images with super wide view in low cost. Image Registration is a key process of Image Mosaic. After studying the correlative technique of Image Registration based on point feature, the thesis does the further research of which selecting matched feature points to calculate the image transformation model.On the basis of the theory of SIFT, the thesis first makes a deep study and experiment about the extraction and registration of SIFT feature points, and then analyzes RANSAC algorithm which used to refine matched feature points to calculate transformation model in Image Registration after getting matched SIFT feature points. The main works of the thesis are as follows:1. Regarding to the deficiencies about the way of selecting points and exclusion condition of candidate models in RANSAC, this thesis revamps the method in the way of selecting points and detecting the candidate models. Experiments show that the improved method takes less time and do better in calculating transformation matrix than original RANSAC algorithm.2. Referring to the idea of RANSAC algorithm, the thesis presents a method of selecting points to calculate transformation matrix based on genetic algorithm. This method avoids the condition that setting the threshold of accepting candidate models by human intervention, thus the method takes some adaptability. Experiments show the horizontal scaling factor and vertical scaling factor of matrix calculated by this method are better than the RANSAC does, and so does the average accuracy.In the end, combined with the framework of Image Mosaic, the thesis carries on the simulation test about image mosaic under the condition of small wide scanner after registering images which using the matrix calculated by the methods mentioned in this thesis. The result of simulation shows the SIFT feature points and the two methods mentioned in this thesis can meet the application requirements.