Research On Recovery Methods For Digital Images

The recovery of digital images is of significance in both theoretical research and practical applications. Form Bayesian inference point of view, this dissertation analyzes the mathematical model of recovery problem, and focuses on some key issues in the topic of image inpainting, image fusion, and recovery-oriented compression applications. The main work and innovations are listed as following:1. According to the analysis on two main PDE-based inpainting methods, that are "macro-inpainting mechanisms" and "micro-inpainting mechanisms" PDE-based, respectively, a combined inpainting method aiming to restore small gap is proposed. Macroscopically, the prior term in proposed energy function mirrors image model. Microscopically, the associated PDEs simulate the generation of image. According to the analysis on texture synthesis-based completion methods, three completion rules are proposed. These rules turn the completion problem into a globally discrete optimization problem which is solved via an EM-like algorithm. The algorithm unifies two key procedures in the completion—patch matching and patch synthesis.2. To fill-in blocks of missing data in wireless image transmission, a POCS-based approach which makes use of the information from both surrounding available blocks and channel estimation is presented. According to the computed edge orientation of a missing block, an adaptive recovery algorithm which combines frequency and spatial domain information is proposed. The proposed approach can restore image edge and complex texture satisfactorily. Improvement in restored image quality and robustness achieves when comparing with RIBMAP.3. According to the analysis on current methods in reconstruction from gradient field, a reconstruction method based on total variation (TV) model is proposed. Editing and stitching applications in image fusion area are introduced. A unified gradient domain stitching approach is presented. We compare our proposed approach with state-of-the-art approaches mathematically and experimentally. Some key issues of gradient domain image fusion are discussed, and an improved method based on boundary optimization is presented. This method can handle both blurring and geometry deformation effectively.4. The compression applications using image recovery technology are studied. We investigate the relation between the compression problem and the recovery problem from mathematical point of view, and present a recovery-oriented compression framework. An image compression scheme guided by this framework is described.