Super-resolution algorithm combining SAE dictionary learning and anchored neighborhood regression

Abstract

To improve the efficiency of the super resolution algorithm based on dictionary learning, a super-resolution algorithm combining sparse autoencoder dictionary learning and anchored neighborhood regression is proposed. The sparse autoencoder with outstanding learning ability is used to learn a dictionary witch has better feature expression ability in the stage of dictionary learning. For the improvement of autoencoder, the mean absolute error principle is taken as the reconstruction error term to improve the accuracy of model error measurement. In the stage of data preprocessing, the whitening technology is used to construct the low redundancy input data to improve the generalization ability of sparse autoencoder dictionary learning model. In the stage of image reconstruction, the dictionary obtained is applied to the super-resolution algorithm based on anchored neighborhood regression to achieve fast real-time reconstruction by reducing the computation of sparse coding. In this study, the proposed super-resolution algorithm combines the advantages of sparse autoencoder model and anchored neighborhood regression, which can not only improve the quality of image reconstruction, but also guarantee the reconstruction speed at the same time. So it has high reconstruction efficiency.