Image registration : algorithms and applications

Image registration is the task of determining positions of corresponding points in two images. A fast and accurate image registration plays an important role in many computer vision and graphics problems. This thesis presents fast and robust pixel and feature-based image registration algorithms and explores new applications in computer vision and graphics. It is divided into two parts. In the first part, we propose a fast and accurate pixel-based image registration algorithm that registers a pair of stereo images to extract the three-dimensional information of the scene. The algorithm finds an appropriate match for each reference image pixel in the target image to compute dense depth maps and works very well in case of repetitive patterns, object boundaries, as well as in occluded and non-textured image regions without increasing computational overhead significantly. The three-dimensional information obtained using pixel-based registration is then used to recognize complex human activities. A sequence of depth maps is represented as a sequence of codewords that are learnt to discriminate between activities. These discriminative sub-sequences are then used to classify and localize an activity in a given test video. To demonstrate the efficiency of these algorithms, we provide experimental results on standard datasets which are publicly available. In the second part, we present a feature-based registration method that registers different image features to find most similar images from very large-scale image datasets. The proposed system retrieves and ranks database images based on their similarity to any given input image. By registering different image features, the retrieval and indexing of the relevant images is improved. We use these similar images to solve an ill-posed computer graphics problem, viz., colorization of grayscale images. The algorithm uses similar color images as the reference images and transfers color information to a given input grayscale image using a graph-based optimization method. The algorithm works at the resolution of superpixels and uses local properties for proper color transfer. Our use of superpixels reduces the complexity of our algorithm significantly and it also empowers the colorizations to exhibit a much higher extent of spatial consistency in the colorization as compared to those using independent pixels. We evaluate these approaches on a wide variety of images. Both qualitative and quantitative analyses have been used to demonstrate the greater effectiveness of the proposed approaches.DOCTOR OF PHILOSOPHY (SCE