Research On The Key Technology Of Image Stitching

Image stitching is a task used to match and register a group of images having acertain overlap area between any two into a reference surface，and then use blendingmethod to form a seamless panoramic image that contains all of the source imagesequence．The main process contains geometric registration and optical registration.Geometric registration is to solve the problem of unifying the two in differentcoordinate system of the image to the same coordinate system, and then it becomes animage. Optical registration is a problem of deciding the pixel gray value of stitchingresulting image. In this paper, we carried out research in the key technologies for thepanoramic image generation, and the main work focuses on the image geometricregistration and optical registration. After comparing many region and feature basedregistration algorithms, we presents a panoramic image registration algorithm based onimproved RANSAC, which selects the classic SIFT feature as the image registrationbasis. In the feature match purification stage, we use an improved RANSAC algorithmto get homography. Finally, the panoramic images are generated by the weightedaverage method to achieved good results. In the optical registration stage, it was foundthat the traditional image blending algorithms tended to fail in the case of largeexposure difference in the source images. For blending images with large exposuredifferences, we present a brightness adjustment algorithm to eliminate the luminancedifference between the mosaic images.The research contents are summarized as follows:1. Presenting a panoramic image mosaic algorithm based on improved RANSAC.Algorithm selects the classic SIFT features to be the image registration basisand gives some improvements on traditional RANSAC algorithm to make itmore suitable for image stitching by using the weighted average method.2. For blending images with large exposure difference, a local brightnessadjustment method is proposed to eliminate the difference in brightnessbetween the images. Experiment results show that the proposed algorithm cansolve the problem in the large exposure difference image mosaic with good result.3. For the problem of brightness inconsistencies in panoramic image thatgenerated for a big scene by pieces of source image sequence, a globalbrightness adjustment method is proposed to eliminate the difference inbrightness of the different regions of the panoramic image. Verified byexperiments, the algorithm can solve the global panorama stitching brightnessuniformity issues.