Improving Bit-Depth Expansion via Context-Aware MMSE Optimization (CAMO)

With the rapid development of consumer devices, display devices at bit depth (BD) of 10 or higher have become increasingly popular, but most mainstream media sources are still at BD of 8. Therefore, ordinary users cannot fully enjoy the advanced equipment, and BD expansion problem is raised recently as the most economical solution to accommodate the gap. Different from existing algorithms which do not optimize a clear objective metric or do not provide optimal solution to the objective, in this paper, we propose a pixel-wise context-aware minimum average mean squared error (MMSE) optimization scheme (CAMO) whose optimal solution can be directly derived to effectively improve the BD expansion performance. Under Bayesian framework, the MMSE objective is decomposed into two separate terms: a likelihood term represents local contexts and a prior term represents image's inherit characteristics. The complicated local contexts are first described via generic Taylor expansion, and then, local derivatives of different orders are estimated to capture non-linear structures. We also investigate different types of images and introduce image-type-wise statistical priors for better performance. To validate the efficiency of the proposed algorithm, we perform extensive experiments on four image datasets, and CAMO achieves promising results compared with state-of-the-art algorithms.