Optimized Regressor Forest for Image Super-Resolution

The goal of image super-resolution is to recover missing high frequency details of an image given single or multiple low-resolution images. It is a well-known ill-posed problem and requires mature prior knowledges or enough examples to restore high-quality high-resolution images. Recently, many methods formulate image super-resolution as a regression problem. Input image patches are classified into pre-trained clusters, and cluster-dependent mapping functions are employed to super-resolve input patches. In this paper, for further improving the reconstructed image quality, an optimized regressor forest framework is proposed, which leverages the discriminative power of random forest. There are three major contributions of the proposed framework. (i) The proposed scheme overturns existing approaches by training the regressors first and learning the way to find the best regressor to avoid quality degradation introduced from the classification outliers. (ii) We propose to employ EM-algorithm to optimize regressors by jointly optimizing the clustering results as well as the regression functions. (iii) In order to find the most appropriate regressor for an input patch at the testing stage, random forest is adopted to accurately classify patches into their best clusters (regressors). The experimental results demonstrate that the proposed method generates high-quality high-resolution images and yields state-of-the-art results.