This paper proposes an example-based Super-Resolution (SR) algorithm of compressed videos in the Discrete Cosine Transform (DCT) domain. Input to the system is a Low-Resolution (LR) compressed video together
with a High-Resolution (HR) still image of similar content. Using a training set of corresponding LR-HR pairs of image patches from the HR still image, high-frequency details are transferred from the HR source to the LR video. The DCT-domain algorithm is much faster than example-based SR in spatial domain6 because of a reduction in search dimensionality, which is a direct result of the compact and uncorrelated DCT representation. Fast searching techniques like tree-structure vector quantization16 and coherence search1 are also key to the improved efficiency. Preliminary results on MJPEG sequence show promising result of the DCT-domain SR synthesis approach.
This paper derives a theoretical limit for image registration and presents an iterative estimator that achieves the limit. The variance of any parametric registration is bounded by the Cramer-Rao bound (CRB). This bound is signal-dependent and is proportional to the variance of input noise. Since most available registration techniques are biased, they are not optimal. The bias, however, can be reduced to practically zero by an iterative gradient-based estimator. In the proximity of a solution, this estimator converges to the CRB with a quadratic rate. Images can be brought close to each other, thus speedup the registration process, by a coarse-to-tne multi-scale registration. The performance of iterative registration is finally shown to significantly increase image resolution from multiple low resolution images under translational motions.
This paper presents a method to predict the limit of possible resolution enhancement given a sequence of low-resolution images. Three important parameters influence the outcome of this limit: the total Point Spread Function (PSF), the Signal-to-Noise Ratio (SNR) and the number of input images. Although a large number of
input images captured by a system with a narrow PSF and a high SNR are desirable, these conditions are often not achievable simultaneously. To improve the SNR, cameras are designed with near optimal quantum efficiency and maximum fill-factor. However, the
latter widens the system PSF, which puts more weight on the deblurring part of a super-resolution (SR) reconstruction algorithm. This paper analyzes the contribution of each input parameters to the SR reconstruction and predicts the best attainable SR factor for given a camera setting. The predicted SR factor agrees well with an edge sharpness measure computed from the reconstructed SR images. A sufficient number of randomly positioned input images to achieve this limit for a given scene can also be derived assuming Gaussian noise and registration errors.