Performance evaluation of multi-frame super-resolution algorithms

Multi-frame super-resolution algorithms aim to increase spatial resolution by fusing information from several low-resolution perspectives of a scene. While a wide array of super-resolution algorithms now exist, the comparative capability of these techniques in practical scenarios has not been adequately explored. In addition, a standard quantitative method for assessing the relative merit of super-resolution algorithms is required. This paper presents a comprehensive practical comparison of existing super-resolution techniques using a shared platform and 4 common greyscale reference images. In total, 13 different super-resolution algorithms are evaluated, and as accurate alignment is critical to the super-resolution process, 6 registration algorithms are also included in the analysis. Pixel-based visual information fidelity (VIFP) is selected from the 12 image quality metrics reviewed as the measure most suited to the appraisal of super-resolved images. Experimental results show that Bayesian super-resolution methods utilizing the simultaneous autoregressive (SAR) prior produce the highest quality images when combined with generalized stochastic Lucas-Kanade optical flow registration.