Single Image Super Resolution Based On Cluster Regression

Single frame image super resolution aims to restore corresponding high resolution image with low resolution input,reconstructing high frequency and details missing in image degradation process.So far the approaches to deal with the SISR problem include: interpolation based,reconstruction based,learning based,and edge priori based methods.In the thesis the learning based super resolution methods are studied and preprocessing methods are discussed.Inspired by several recent learning based super resolution methods,a novel SISR method using joint constraints by learning based gradient and high frequency estimation called Single Image Super Resolution Through Multi Extreme Learning Machine Regressor Fusion is proposed,the main work are summarized as follows: 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 calculated based on the distance of the cluster centers with patch isotropic characteristics.Then,the estimated HR image gradient and high frequency are used 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 has competitive SR quality compared with the other state-of-the-art SISR methods.Considering compression artifacts and noise exist in images widely,appropriate processing methods are studied to ensure the quality of image super resolution in the thesis.To handle compression artifacts in compressed images,the necessity of handling compression artifacts properly for image super resolution is analysed in the thesis and an algorithm framework for compressed image super resolution reconstruction is proposed,sample clustering and nonlinear regression for reducing image compression artifacts is adopted as processing method,which composes the algorithm framework with the following super resolution step.To handle image noise,a Nonlocal means method based on multichannel joint estimation for color image denosing is proposed,including two steps as color channel combination filtering and color channel fusion filtering: In color channel combination filtering step,the noisy color image is denoised by the classical NLMC,from which the pre denoised image is obtained as the input of color channel fusion filtering step;In color channel fusion filtering step,the pre denoised image is denoised once more by generalized multichannel NLM,and the similarity between the high frequency components of the pre denoised image's RGB channels is used in the denosing process at the same time.The experimental results demonstrate that the proposed method produces competitive results in both quantitative and visual comparisons with other classical color image denosing algorithms,and the reconstruction quality of noisy color image can be significantly improved by the proposed method as preprocessing step.