A Review of Intrinsic Image Decomposition

Intrinsic image decomposition (IID) aims to extract intrinsic components from natural images, namely shading and reflectance, which is widely applied in computer vision and image processing tasks for the better understanding of the objects and scenes. However, IID is a severely ill-posed problem, extra constraints should be applied to address the decomposition problem. In this paper, we reviewed recent papers about intrinsic image decomposition. The extra cues from lighting, shape, color and textures are used to constrain the ill-posed problem. These papers can be mainly classified into two categories: physics-based methods and learning-based methods. Physics-based methods mainly introduced extra constraints into an optimization framework to solve the IID problem. While learning-based methods mainly depend on an encoder-decoder network with the explicit constraint from the labeled intrinsic images. Additionally, extra priors are introduced into the network by the form of network structure or specific feature. In order to make the IID problem more practical, researchers may explore how to achieve fast and accurate image decomposition with lower computing resources through optimization algorithms and model design.