Super-resolution imaging : Markov chain Monte Carlo and state-space approaches

Abstract

List of Figures 1.1 A conceptual framework of super-resolution imaging [1-19] 2 1.2 Various versions of the Text image: (a) original image (50 x 50); (b) enlarged image (200 x 200) by applying pixel duplication on the image (a); (c) enlarged image (200 x 200) by applying super-resolution algorithm on the image (a) 3 1.3 An overview of the proposed stochastic SR image approach using the Markov chain Monte Carlo techqnue 4 1.4 The proposed three-equation-based state-space representation for representing the relationships among the low-resolution frames and highresolution frames 6 2.1 The relationship between the original high-resolution image and the observed low-resolution images. The observed low-resolution images are the warped, blurred, down-sampled and noisy version of the original high-resolution image 11 2.2 The registration-interpolation-restoration-based approach for superresolution image reconstruction 17 2.3 A conceptual framework of learning-based SR image approach. ... 22 2.4 The sliding-window-based super-resolution video scheme 27 2.5 The simultaneous multiple high-resolution frames reconstruction scheme. 28 vii ATTENTION: The Singapore Copyright Act applies to the use of this document. Nanyang Technological University Library LIST OF FIG URES LIST OF FIG URES 3.1 Various reconstructed results of Boat image: (a) the original image; (b) the simulated quarter-sized low-resolution image; (c) the bi-cubic spline interpolation approach, followed by deblurring (PSNR = 25.80 dB); (e) Vandewalle et ai's algorithm [20] (PSNR = 25.97 dB); (g) Pham et ai's algorithm [21] (PSNR = 26.61 dB); (i) Farsiu et a/.'s algorithm [22] (PSNR = 27.45 dB); (k) the proposed algorithm with the automatically-updated hyperparameter (PSNR = 28.02 dB); images (d), (f), (h), (j) and (1) are the absolute difference between image (a) and images (c), (e), (g), (i) and (k), respectively. Furthermore, the intensity values of these absolute difference images are scaled up three times for demonstration 49 3.2 Various reconstructed results of Text image: (a) the original image; (b) the simulated quarter-sized low-resolution image; (c) the bi-cubic spline interpolation approach, followed by deblurring (PSNR = 17.62 dB); (e) Vandewalle et a/.'s algorithm [20] (PSNR = 18.32 dB); (g) Pham et a/.'s algorithm [21] (PSNR = 17.74 dB); (i) Farsiu et aVs algorithm [22] (PSNR = 19.29 dB); (k) the proposed algorithm with the automatically-updated hyperparameter (PSNR = 20.17 dB); images (d), (f), (h), (j) and (1) are the absolute difference between image (a) and images (c), (e), (g), (i) and (k), respectively 50 viii ATTENTION: The Singapore Copyright Act applies to the use of this document. Nanyang Technological University Library 65 3.10 The image sample, which is generated at each sample generation step of the proposed MCMC SR approach, is enhanced by adding a positive scaling factor j3 that is multiplied with a linear high-pass filtered version of the image sample 71 3.11 Three test images used in the experiments: (a) 256 x 256 Airplane; (b) 256 x...