Date of Award
Doctor of Philosophy (PhD)
Chung Hao Chen
Gene J. Hou
Single image super-resolution (SR) is a technique that generates a high- resolution image from a single low-resolution image [1,2,10,11]. Single image super- resolution can be generally classified into two groups: example-based and self-similarity based SR algorithms. The performance of the example-based SR algorithm depends on the similarity between testing data and the database. Usually, a large database is needed for better performance in general. This would result in heavy computational cost. The self-similarity based SR algorithm can generate a high-resolution (HR) image with sharper edges and fewer ringing artifacts if there is sufficient recurrence within or across scales of the same image [10, 11], but it is hard to generate HR details for an image region with fine texture.
Based on the limitation of each type of SR algorithm, we propose to combine these two types of algorithms. We segment each image into regions based on image content, and choose the appropriate SR algorithm to recover the HR image for each region based on the texture feature. Our experimental results show that our proposed method takes advantage of each SR algorithm and can produce natural looking results with sharp edges, while suppressing ringing artifacts. We compute PSNR to qualitatively evaluate the SR results, and our proposed method outperforms the self-similarity based or example-based SR algorithm with higher PSNR (+0.1dB).
Zhang, Min. "Region-Based Approach for Single Image Super-Resolution" (2016). Doctor of Philosophy (PhD), dissertation, Electrical/Computer Engineering, Old Dominion University, https://digitalcommons.odu.edu/ece_etds/12