Single Image Super Resolution Through Multi Extreme Learning Machine Regressor Fusion

Abstract

Single image super resolution (SISR) aims to generate a high resolution (HR) image based on a given low resolution (LR) input. The edge priori based SISR methods tend to estimate HR image by edge-preserving constraint. In this paper, a novel learning based SISR method is proposed to reconstruct HR image by using joint HR gradient field and high frequency constraint. In the training phase, interpolated training LR patches with similar structure are partitioned into the same cluster by K-means clustering, and the Extreme Learning Machine (ELM) are used to get gradient and high frequency regressors in each cluster by training LR/HR patch pairs. In the prediction phase, multi-ELM regressor fusion strategy is used to estimate more accurate gradient and high frequency data, in which the fusion weights are based on the distance of the cluster centers with patch isotropic characteristics. Then, the estimated HR image gradient and high frequency are regarded as a joint constraints priori to reconstruct HR image. Experimental results demonstrate that the proposed method achieves better estimating accuracy of gradient and high frequency and have competitive SR quality compared with the other state-of-the-art SISR methods.