Feature preserving super-resolution use of LBP and DWT

In this paper, we propose a novel technique for feature preserving image super-resolution. Given a test image and a training database consisting of low resolution images and their high resolution versions, we obtain the super-resolution by learning the details from the high resolution training images. The image features such as edges, corners, and curves carry crucial information in many imaging applications. We exploit the fact that the local geometry of these features in the low resolution image is similar to that of their high resolution counterparts. We model these features using local binary patterns that best capture underlying geometric information and obtain super-resolution of the image by learning discrete wavelet coefficients of the high resolution features present in the training images. The experiments are conducted on real world images and results are compared with recently proposed super-resolution approaches. Experiments show that the proposed approach perform better in both qualitative and quantitative evaluation. The approach is promising for real time application as it is non-iterative.