Study On Image Registration And Mosaicing Method

With the rapid development of digital imaging technology, digital imaging equipments have been widely used. However, in some situations, the digital image and video can not met the demands because of the limited field of view. This paper studies the digital image mosaicing technology to generate large field-of-view panorama image from image sequence or video.In general, image mosaicing process consists of the following steps: image acquisition, image registration, image re-projection, and blending processing. Image registration is the foundation of image mosaicing. In this paper, the principle of geometric image registration and photometric registration are reviewed. The transformation optimization image registration method and FFT-based phrase correlation image registration method are studied respectively. A new automatic multi-scale multi-resolution image registration algorithm is proposed. The input images are decomposed into gradient pyramids. The sub-images in each pyramid layer are thresholded adaptively and then fused to find out feature points. The quantity of feature points are selected proportionally according to the variation of gradient distribution for decreasing computing cost. The images in coarse layers are registered based on feature points. The registration starts from the coarsest layer and the result is up-sampled as the initial estimation for the finer layer. Different transformation models are used for each layer, from simple to complex as the image resolution increases. At the original image size layer, the registration result is further tuned by Levenberg-Marquardt nonlinear optimization and bilinear interpolation method on image pixel gray level. The experiment results indicate that the proposed algorithm is effective.The re-projection model can express the panorama image in different ways. Four different kinds of re-projection model are studied. These models are planar manifold, cylindrical manifold, spherical manifold, and adaptive manifold model. Based on the comparison results, the adaptive manifold is used in this paper as the re-projection model to cerate panorama from image sequence and video.A good blending function can erase the artifacts in mosaics. In this paper, several kinds of blending function are comparatively analyzed. The experiment result shows that the nearest image center blending function can produce relatively good blending.