Structure and Texture-Aware Image Decomposition via Deep Optimization

  Structure-texture image decomposition is a fundamental and challenging topic in computational graphics and image processing. In this paper, we introduce a structure-aware measure, as well as a texture-aware one, to facilitate the structure-texture decomposition (STD) of images. Instead of the edge strengths and spatial scales that have been widely-used in previous STD researches, the proposed two measures differentiate image textures from image structures via distinctive motivations.

  To benefit the STD problem, both the measures are further involved in an objective function by weighting its regularization terms. The objective function is optimized by training a light-weight neural network so that we do not require to design specific optimizers for different functional spaces in the objective function. Moreover, the network is trained only by the input image itself, rather than a larger number of external samples. The experimental results demonstrate that our method can separate image structure and texture in a better way and result in shaper edges in the structural component, in comparison with some state-of-the-art methods.

Structure & Texture Measure

Objective Function

Network Architecture

Results