Research On Fully Automatic Image-stitching Technologies For Plane-like Scenes

Image stitching is to combine and blend a sequence of overlapped images togenerate a panoramic image with wider view and higher resolution, which can helpviewers achieve a more general and direct understanding of the scenes. Nowadays,image stitching has been widely used in remote sensing, military affairs, digitalentertainment, etc. and became important research topic in the field of computer vision,computer graphics, and image processing. Image stitching for plane-like scenes is alsocommonly used in many applications, such as generation of satellite maps, scanning oflarge files, aerial monitoring and surveillance. However, image stitching for plane-likescenes has its own particularities and requirements, so traditional image stitching is notfit for plane-like scenes. Moreover, various existing methods designed for plane-likescenes also have respective drawbacks. Therefore this paper focuses on image-stitchingtechnologies for plane-like scenesMain works of this paper include (1) Analyzing and comparing many basictheories, crucial technologies and optimizations in the field of image stitching.(2)Based on the framework of motion-based stitching, a novel overall structure is proposed,and techniques in stereo vision are induced. Then a integrated and practical "automaticstitching system for images of plane-like scenes"--ASIPS is proposed andimplemented.(3) A fresh new module called "dominant plane extraction and warping toquasi-perpendicular image" is added to the system.(4) The last step of stitching isre-designed as "incremental alignment and image blending" module. The formermodule firstly estimates the pose of camera, then performs sparse reconstruction for thescene to estimate the scene structure. A dominant plane is then extracted, and with itsdirection images are warped to "quasi-perpendicular view images". The latter moduleadopts similarity transform model to incrementally align warped images, and uses thenovel "simplified multi-band image blending" algorithm to combine aligned images intoa panorama.This paper also put forward some standards to evaluate the stitched panorama,including qualitative standards for evaluating the visual attraction and quantitativestandards for evaluating the correspondence between image and scene. Plenty ofexperiments are carried out in both simulative scenes and real scenes to test the visualattraction and the correspondence, and those proposed standards are used to objectively evaluate the results. Results of the experiments prove that ASIPS has followingadvantages: satisfactory visual effects, accurate correspondence between image andscene (and the output panoramas can be perceived as taken with the directionperpendicular to the scene), ease of image acquisition, strong robustness, highadaptability and full automation. Following works will solve the problem that the visualattraction of ASIPS is downgraded when there is heavy parallax.