The Large FOV Image Mosaicing: Theory And Applications

With the development of digital recording equipment, we hope that the images or videos we get become more interesting and useful, therefore, the researchers begin to find out various applications on digital images and digital videos. This paper mainly discusses various expression ways of image sequences and video, which break through the physical restriction of equipment. We make use of large FOV image mosaicing to construct panoramic image, from which we can make panoramic QTVR.Generally, this paper discusses three main steps of image mosaicing: image registration, reprojection manifold, and blending processing.First, image registration is the base of image mosaicing, who is also the base of other image processing, such as image fusion. In the paper, we separatively discuss image geometric registration and photometric registration. In the domain of geometric registration, we separatively discuss the direct image registration and feature-based image registration, and make compare with each other. We adopt the frequency-domain-based affine model and optical flow equation. First, we compute the larger spatial transformation. Then, we refine the parameters using projective transformation. It is proved that this method can produce good result and fast speed. Moreover, in the domain of photometric registration, we adopt a photometric model to correct the poor exposure images.Then, we list four different kinds of reprojection model: planar manifold, cylindrical manifold, spherical manifold, and manifold projection. Those models express the panoramic image in different ways. In our system, we adopt planar manifold as our reprojection model of panorama, and adopt cylindrical manifold as our panoramic QTVR reprojection model. For round scenes, the cylindrical manifold can actually restore the original scene style.Without blending processing, there must be cracks in the mosaics. So we must find out a good blending function to erase those cracks. Here, we list kinds of blending function, and make compare with each other. Finally, we choose the nearest image center blending function as our blending function, which produces good result.Furthermore, for constructing a large field of view panoramic image, we should loop aroundto shoot some parts of one scene, which may produce wrong displacement of spatial continuous but time discontinuous images. This global discontinuous is produced by homography multiplication, so we must do some optimal processing. At the same time, we eliminate the image blur by motion objects using optimal path finding.At last, we make use of the panoramic image to re-project to the cylindrical panoramic QTVR, which expands the application domain of image mosaicing, and need higher requirement of image mosaicing.