Development of Efficient Algorithms for Seamless Image Mosaicing

Abstract

This thesis investigates the problem of seamless composition from multiple input images to produce a visually pleasant image mosaic. This area is also referred as image stitching or image compositing, which has the potential of increasing the field-of-view (FOV), removing visual artifacts, editing images, smoothing seam, and improving the quality of the mosaiced images. The proposed work has diverse applications in the area of panorama generation,image editing, special visual effects, and multimedia image compositing.
Two interrelated forms of image mosaicing are investigated. Firstly, novel algorithms for image compositing are proposed. Image compositing is a process of creation of completely new images, by combining two or more image regions seamlessly from non-overlapping images. For seamless composition, new blending algorithms are proposed. Further, to make the mask generation process automatic, saliency based object detection methods are developed. The details of the propositions and the implementation are discussed for the complete image compositing framework.
Secondly, new methods for image stitching, which is the process of aligning a series of images with some overlapping parts to generate an image with wider FOV and no visible
seam are proposed. In particular, a discrete cosine transform (DCT) based stitching and a feature based stitching approach are developed for panoramic mosaicing. One of the important steps prior to blending of images is image alignment. A blur invariant image alignment method using DCT-based phase correlation is proposed to get image mosaics. To deal with illumination or exposure difference between the input images, a feature-based registration with integration to graph-cut approach is presented.
Throughout this thesis work, the performance of the algorithms is evaluated using both synthetic and publicly available dataset. The proposed algorithms enable the
automatic generation of visually pleasant mosaics for both compositing as well as stitching. Several applications are demonstrated for the proposed mosaicing algorithms such as
image compositing for visual effects, creation of new images without user/with least user intervention for multimedia applications and panorama generation with wide FOV.