Single image super resolution using dictionary learning and sparse coding with multi-scale and multi-directional Gabor feature representation

Abstract

Recently, sparse representation theory has shown the state-of-art performance in image super resolution (SR) field. The sparse representation based single image SR methods learn dictionaries to discover the co-occurence relationship between low-resolution (LR) and high-resolution (HR) image feature spaces to generate satisfactory SR images. However, most existing learning based methods focus on Gradient and Laplacian filters to achieve accurate feature representation in SR task. In this paper, a novel multi-scale and multi-directional feature descriptor approach is proposed to improve SR quality. In the proposed approach, Gabor filter is first used as feature representation in learning based super resolution to extract image features at different scales and orientations. Therefore, the difficulty of capturing the complex local structures in all scale and directions using traditional 1-D filters is resolved by the proposed Gabor filter approach. Then the efficient mapping between LR and HR images is achieved by searching the sparse representation coefficients over the LR and HR dictionaries in dictionary learning phase. Finally, the learned relationship is applied to validation LR input in order to achieve accurate SR image in reconstruction phase. The experimental results show that the proposed method outperforms other state-of-art methods in terms of both quantity and quality.