Research on Automatic Stitching
|School||Zhejiang University of Technology|
|Course||Applied Computer Technology|
|Keywords||Image Stitching Image Registration Image Blending SIFT Error Probability Density|
Image mosaic means to stitch a series of overlapped images into a bigger imagewith super-resolution and large angle view, which includes two key steps, registrationand blending. Registration is mainly related to the geometry adjustment of the imageposition and Integration is mainly related to the consistency adjustment of color andbrightness of the image. Image mosaic includes2D and3D image mosaic.2D imagemosaic aims to build high resolution, wide Angle image, while3D images mosaicpurposes to obtain complete3D information of the object.With the development of science and technology, Image mosaic has become avery important technology and has extensive and practical applications in the field ofmodern information processing, such as space exploration, medical, military, videoretrieval and transmission. However, image mosaic still has a bottleneck ofapplication in many fields. At present, a lot of important computer vision research isbased on the hypothesis that the stitching problems have been resolved. So theresearch of image mosaic is still of great value.At first, this paper introduces the background and significance of researching theimage mosaic by summarizing and classifying the present stitching methods. For thepreparation of the following research, the paper also describes the coordinate systemand some image transformation models in stitching.This paper mainly researches the2D image stitching and3D image stitching. Theregistration is the most important step of image stitching by using local and globalway. The research on2D image registration based on the SIFT feature method showsthat as a local registration, it is difficult to stitch images with single scene and toextract enough feature points for stitching.Thus the paper proposes a2D image registration method based on a global way.This method achieves image registration by constantly optimizing errors in the overlapped image area to adjust the geometry transform parameters between images.The method is more stable because it uses all image information.For3D image mosaic, the paper introduces a registration method based onprobability density. This method achieves image registration by establishing mixedGaussian probability density function and using conjugate gradient algorithms toconstantly optimizing difference value between the mixed Gaussian probabilitydensity function. The method also uses all3D image information and is a globalimage registration method.But the image color and intensity have nonlinear difference because ofenvironment changing and collecting device, which can influence the image stitching.So referring to the existing image blending methods, this paper uses the simple linearblending method and multi-resolution spline blending to complete image blending.Many experiments validates that the image registration based on globalinformation in the paper is sufficient to adjust geometry position between images. Themethod can also realize the consistency adjustment of the image color deviation andintensity inconformity between images. In addition, it is more stable.