Research On Panoramic Image Stitching Based On Improved Method Of Image Registration

In recent years, panoramic images are widely used in the areas of virtual reality,computer simulation, and computer vision, etc. In addition, a series of applications based onpanoramic images also have been developed, such as street view, immersion multimediadisplay and panoramic video surveillance, etc. Because of the photographic equipment whichcan obtain panoramic image directly having some blindsides, such as high price and lowpenetration rate, we usually use camera to obtain the original images, and get the panoramicimage by image stitching technology. Panoramic image stitching technology can make somedispersed images which describe different scenes joined together into a panoramic image thatcan describe scenes comprehensively, by detecting, recognizing and progressing on overlapregions between images. Thus, the methods that can improve the quality and efficiency ofpanoramic image stitching have become a hot research topic.This paper introduces the pipeline of panoramic image stitching system in detail,including image acquisition, image preprocessing, image registration, image transformationand image fusion. Image registration is the core process of this pipeline, whose efficiency candeeply affect the speed of image stitching. What’s more, inaccurate image registration maycause misallocation of panoramic image. Image registration algorithm based on features hasstrong robustness, better able to cope with scale transformation, scaling transformation andlighting differences between images, so it has become the commonly used image registrationalgorithm, and its pipeline includes feature detection, feature description, feature matchingand the features fitting, in which feature detection and feature description are core processes.In practice, the original images for panoramic image stitching can be obtained in avariety of application scenes. The existing methods of image registration usually use singlealgorithm of feature detection and feature description to deal with image registration fordifferent types of images, and the efficiency of mismatches eliminating should be raised.Considering this problem, we firstly conclude and analysis the reality of different applicationscenes, and then propose a new algorithm of panoramic image stitching based on thecombination model of feature matching. In this model, we firstly use binary algorithm of feature detection and description with high speed for the first level image registration, andthen take feature fitting with a restricted strictly fault-tolerant threshold, so that the imageregistration problems of most application scenes can be solved. If the fitting fails, we use thealgorithm of feature detection and description based on multi-scale with strong robustness forthe second level image registration, and also use the improve algorithm proposed by thispaper to accomplish mismatches eliminating, so that the image registration problems of somespecial application scenes, such as scenes including huge amount of similar structures, can besolved. Because of the lack of public experimental image database for panoramic imagestitching, this paper builds the image database for experiments, including variety ofapplication scenes. Then we choose several algorithms to take contrast experiments. Theresults of experiments show that the algorithm proposed by this paper achieves high accuracyand speed of image registration.Lots of mismatches may occur when the images include huge amount of similarstructures, and the high mismatch rate may reduce the speed, success rate and accuracy offeature fitting which will affect the success rate, quality and efficiency of image stitching. Soit is necessary to improve the speed and quality of the existing mismatches eliminatingalgorithms. Considering this problem, this paper proposes improved mismatches eliminatingalgorithm based on color information and binary global feature. Firstly, we use extendedSURF feature descriptor based on HSI color model for feature description, so that mismatcheswhich have big differences in color image information around the feature points’neighborhood with small range will be eliminated mostly. After feature matching, we use thealgorithm based on binary global feature to eliminate mismatches, so that mismatches whichhave big differences in the relative position of global region will be eliminated mostly. Binarystring is used to implement binary global feature’s rapid calculation, and the integral image isintroduced, so as to improve the computing speed of the algorithm. Then we choose severalalgorithms to take contrast experiments, and the experimental results show that the accuracyof matches is increased by about5percent, and the speed of the algorithm proposed by thispaper is about1.7times faster than the existing algorithms.