Unrolled Optimization with Deep Priors for Intrinsic Image Decomposition

Intrinsic image decomposition is a challenging task, which aims at separating an image into reflectance and shading layers. Traditionally, strong hand-crafted priors such as reflectance sparsity, shading smoothness and depth information, have been used to solve this long-standing ill-posed problem including two variables. Recent researches lay emphasis on the deep neural networks which need to be specific design. To overcome these limitations, we develop a novel unrolled optimization model for intrinsic image decomposition, which incorporate deep priors from the optimization perspective in a more skillful way, rather than directly design the specific network or introduce hand-crafted and human annotation priors. Extensive experimental results illustrate the excellent performance of our method compared with other state-of-the-art methods and we successfully carry out the proposed algorithm for the application based on image decomposition (e.g. low-light image enhancement).