Super-resolution using DCT based learning with LBP as feature model

In this paper, we propose a novel learning based technique for feature preserving super-resolution of a low resolution observation. The local geometry of an image is conveyed by image features such as edges, corners and curves. We encode these features with local binary pattern operator. The missing high resolution features of the low resolution observation are learnt in the form of discrete cosine transform coefficients from high resolution images in the training database. Experiments are conducted on real world natural images and results are compared with the standard interpolation techniques. Both the qualitative and quantitative comparisons show the effectiveness of the proposed approach.