Sparse coding for super-resolution via K-means classification

Aoran Xiao, Zhenfeng Shao, Zhongyuan Wang

Abstract: Super-resolution techniques reconstructing a higher resolution image from one or multiple low-resolution images are helpful to visual recognition under the scenarios of insufficient acquisition resolution. Due to the limited wireless network transmission bandwidth or mobile device processing capacity, image resolution in mobile phones and other mobile devices is not as high as expected, which restricts the clarity of the image display. This kind of resolution-improvement technique shows a great value for the mobile business. wing to the graceful theoretical foundation, image super-resolution methods via sparse coding have gained much popularity since it was proposed in 2008. However, the underlying single sparse dictionary in the existing methods is difficult to adapt to complex and diversified image contents so that the promotion of representation precision heavily depends on the growth of dictionary volume. In this paper, we propose to learn a set of coupled-dictionaries corresponding to unequal content complexity by using K-means classification, and then adaptively choose an appropriate one for each input low-resolution patch during super-resolution reconstruction. The experimental results on public dataset show that our approach outperforms the classic sparse coding based methods in terms of visual effects and PSNR metrics.