Comparison of reconstruction and example-based Super-Resolution

Abstract

By comparing two classes of Super-Resolution (SR) namely Example-Based Super-Resolution (EBSR) and Reconstruction-Based Super-Resolution (RBSR), we investigate two points: Firstly, which SR technique EBSR or RBSR will produce SR image that preserves Structure Similarity (SSIM) to the original image? Secondly, which SR technique will produce SR image that is more appealing to human eyes? For resultant SR image, EBSR predicts the relation between high and low frequencies in an image, whereas RBSR algorithms rely on a sequence of frames. From the experimental results on test images, we find that compared to RBSR, EBSR preserves the structure of the original image. Knowing this capability is important for detection and recognition systems. In terms of visual appearance, RBSR is preferred except when there are large motions in consecutive frames. Moreover, the aliasing artifacts cannot be removed by EBSR algorithms.