O uso do SSIM como critério de convergência em abordagens Bayesianas variacionais de superresolução multiframe

Abstract

Multiframe Superresolution is a technique that generates one high resolution image from several lower resolution images. Different SR methods attempt to implement superresolution successfully. Some of them use bayesian approaches and are based on 2-norm or 1-norm. The estimated high resolution image is obtained, in general, by running an iterative algorithm that usually uses MSE (Mean Squared Error) as its convergence criteria. However, MSE does not consider the image’s structural characteristics that are perceived by the human eye. The structural similarity index method (SSIM), on the other hand, uses such characteristics, and luminance and contrast as well, to quantify the differences between the image estimated by superresolution and the original high resolution one. Therefore, in this work, we propose the use of SSIM as an error metrics to lead the iterative adjustment process to obtain the high resolution image, in order to explore in a better way, the gains obtained from a priori functions that preserve edges and structural details. In this regard, we have compared results from different variational Bayesian methods by changing the convergence criteria from MSE to SSIM. A set of experiments were proposed and the results showed that, in addition to the improvement of the quality image assessment, there was an improvement on the execution time of the algorithms.