|About the Book|
Image enhancement refers to image processing algorithms that take a degraded image or a sequence of images as input and produce an image or a sequence of images of better quality. The definitions of degraded as well as good quality images areMoreImage enhancement refers to image processing algorithms that take a degraded image or a sequence of images as input and produce an image or a sequence of images of better quality. The definitions of degraded as well as good quality images are extremely subjective and mostly depend on the application. The degraded image can be blurred, noisy, of lower pixel resolution or even be corrupted by compression algorithms. Based on the subjective definition of the degraded image the algorithms as well as the end results may greatly vary.-Most enhancement problems are computationally intensive and are often ill-posed. By ill-posed we mean that the original de-gradation process, which we are trying to reverse, is a many-to-one mapping, and hence, unique inverse does not exist. Video applications have additional space complexity which makes it necessary to process them in an online manner.-In this work we have attempted to develop algorithms that can reconstruct the images after they have undergone varied amounts and kinds of degradations. We unify the driving motivation of the different problems addressed in this thesis under the common name of image enhancement. We have used deblurring, superresolution, and enhancement exchangeably and have tried to develop algorithms which are generic and can be applied for applications across the different definitions with minor modifications.-We build up an example-based learning framework for image enhancement for single frames. The contribution of this work is identifying the nonparametric approach to solve the problem which is more suitable for enhancement applications. We extend the single frame algorithm and introduce a similar method to handle multiple frames, with known flow fields. Another contribution of this work is the recognition and restoration loop, which works as a closed loop feedback system. We introduce a generic multilayer graphical model that unifies low-level vision tasks, such as restoration, with high-level vision tasks, such as recognition in a cooperative framework. The multiple layers of the graphical model are designed to aid each other such that improvement in one layer entails improvement in the other.-To address more realistic multiframe problems, we introduce an ordinal regression based framework to determine subpixel shifts between image frames. We take a supervised approach where the availability of good quality training data is required to learn a ranker, which can differentiate between different sub-pixel shifts. We further extend the single frame example-based learning framework to remove compression artifacts from video frames.-In the second part of this dissertation we propose a robust large scale system which tries to enhance image patches based on a local manifold based nearest neighbor search. We introduce locality sensitive hashing schemes to speed up the search and demonstrate a prototype for one of the first implementations of an on-demand image enhancer, which can be deployed as a web application. We take on more practical applications such as text image enhancement captured by point-and-shoot cameras.-We wind up the work presented in this dissertation by presenting a subjective study of changes in human interpretation of visual stimuli as a function of image degradation. We stumble on the well known but less understood fact that enhancement does enhance human interpretation, though the quantification of such behavior has still remained an open question.