03-08-2016, 09:17 AM
Image denoising Inspired by the symmetry of natural objects, we propose geometric resampling technique and develop a patch-based nonlocal distributed representation (NDR) for image signals. Using NDR, we show how to exploit the nonlinear manifold constraint by locally linear transforms (LLT) in the patch space (an extension of locally linear embedding) and probabilistically combine the inference results under a Bayesian framework. In addition to a theoretical explanation of recently proposed BM3DDFT algorithm [2], empirical Bayesian perspective enables us to achieve even better denoising results for images of objects satisfying certain symmetry properties.