The Research On Panorama Composition Method Based On Feature Point

Panorama composition technology not only plays an important role in virtual reality,but also gains widespread applications such as medical image processing,remote sensing image processing,video compression, video retrieval,the scope of view and resolution improvement,and collection and displays of military images,especially of infrared images.This thesis focused the research on characteristic point-based panorama composition method based on the analysis and summarization to the existing panorama composition technology,and mainly have done the following four aspects of work:①Performing analysis and comparison to the current several common image projection methods,and choosing cylindrical projection method as the finally image projection method for panorama composition of these thesis,for the reason of its easy images acquisition and the better performances of algorithm implementation.②Proposing an improved Harris corner detection algorithm.As the accuracy of corner detection is the key to image registration,this thesis constructed Harris function using the energy changes of local window in the original Harris algorithm,and presented an improved Harris corner detection algorithm to solve the inconvenience of the Harris function threshold value selection.This corner detection algorithm performs curve fitting to the Harris function from x and y directions respectively,and then finds the peak points of fitting coincidence,putting them as corners, thus can improve the accuracy of corner detection.③Presenting a new image registration method that combines Normalized Cross-Correlation(simply for NCC) corner matching algorithm with the improved RANdom Sample Consensus(simply for RANSAC) algorithm.After detecting precise corners,based on the study and analysis of existing image registration algorithms,this thesis proposed a fast and effective image registration algorithm.It makes full use of the detected corner information of two images,and uses NCC algorithm to perform rough corner registration.Meanwhile,to increase the number of correct matched corner pairs,it uses the improved RANSAC algorithm to delete mismatched corners for two times.Finally, this algorithm performs Levenberg-Marquardt nonlinear optimization for the parameters of affine transformation model to further reduce the image registration errors.④Proposing a panorama stitching and fusion algorithm.This thesis chose a smaller total error image stitching mode of mosaic image to mosaic image,and selected the linear transition of overlap region method to process image after stitching,all these based on the study and analysis to the current panorama stitching mode and image fusion algorithm.The experiment results show that it can implement good image composition effects.