Super-resolution via supervised classification and independent dictionary training

Super-resolution (SR) reconstruction plays an important role in recovering the image details and improving the visual perception. In this paper, we propose a new and effective method based on the idea of classification reconstruction and independent dictionary training. Firstly, we extract some geometric features of images and design a new supervised classification method, which uses the decision tree to guarantee a better classification result. Secondly, the coefficients of the high-resolution (HR) and low-resolution (LR) patches are not equal strictly in fact, which enlighten us to train the HR and LR dictionaries independently. And then mapping matrices are learned to map LR coefficients into HR coefficients, which can not only help us improve reconstruction quality, but also just perform sparse coding one time in the reconstruction stage. At last, we enforce a global optimization on the initial reconstruction HR image based on the non-local means and the auto-regressive model. The experiments show that the method we proposed works better than other classic state-of-the-art methods.