Super-Resolution Model Using Multi Directional Structure Integrity Prior

In this work, we propose a super-resolution reconstruction (SRR) method for estimation of fine structural features in the high resolution (HR) image, from gradients computed along different directions. To preserve and sharpen the fine structural details in the HR image, the edge details along different directions are estimated using multi-directional total variation (TV). The drawback of finite difference approximation of the directional gradients is alleviated by representing the directional features using sparse linear combination of adaptive over-complete bases. The sparse features representative of different directions are then combined using choose-max rule, and introduced as a learned prior in the regularization term. The resulting super-resolution (SR) reconstruction problem with proposed multi-directional structure integrity (MDSI) prior is solved using alternating minimization. The quantitative and qualitative results in terms of peak signal-to-noise (PSNR), structure similarity index measure (SSIM) and signal-to-noise (SNR) demonstrate the superiority of the proposed approach in comparison with low rank total variation (LRTV), joint regularization based super-resolution (JRSR), k-singular value decomposition (K-SVD), block K-SVD (BK-SVD) and enhanced non-local total variation (ENLTV) methods.