Research On Panorama Stitching Based On Speeded Up Robust Features

With the increasement of application requirements, image stitching technology hasbecome an important research subject in the field of virtual reality, computer vision, computergraphics, video processing and so on.This paper gives a systematic exposition of the panorama stitching technology based onits two key techniques: image registration and image fusion. Some improvements have beenproposed based on the current algorithms.In image registration, this paper elaborates the speeded up robust features (SURF)detection algorithm with the scale-invariant feature transform detection algorithm as acompare. In order to get a higher accuracy in the feature matching proceeding, this papermakes use of the feature pairs'symmetry property. The matching process is done by thefollowing steps: firstly, coarsely matched features are found by using the kd tree as a searchstructure and the nearest neighbor as a strategy; secondly, a vast majority of the falsematching pairs are discarded by judging whether they have symmetry or not; finally, atraditional random sampling verification (RANSAC) algorithm is implemented for a furtherfalse feature pairs discard and the image transformation parameters'calculation.In image fusion, this paper uses the multi-resolution spline technique combined withgraph cut-based seam search technique to achieve a smooth transition of the overlappingareas and ghost-free effects caused by moving objects in the input images. It also improvesthe initial weighted direction graph's calculation which is based on the difference betweenpixel intensities. By considering not only the difference between pixel intensities but alsoimage gradient information, a more reasonable image cutting result can be achieved with theimproved weight calculation method. In addition, If the input images have a great exposuredifference, an unnatural color transition phenomenon will exist in the stitched image and theblock-based exposure compensation method can be implemented to fix it.Panorama stitching experiments have been done using algorithms described in this paper.The experimental data confirmes the feasibility and good result of the improved methods.