An Unrolling-based Approach for Structure-Texture Image Decomposition
Please login to view abstract download link
The problem of structure-texture image decomposition has been widely investigated over the past few decades and has found successful applications in tasks such as image abstraction, HDR tone mapping and super-resolution. Classical approaches rely on a variational formulation of the problem and minimize the sum of different energy terms, each encoding specific properties of each component. Recently, motivated by progress in deep learning techniques, various data-driven approaches have been presented exploiting unsupervised learning methods and Plug-and-Play frameworks. In this work, we leverage the unrolling technique to tackle the structure-texture image decomposition problem and we present a neural network LPR-NET based on the unfolding of the Low Patch Rank model. This approach enables automatic parameter learning from data in a supervised fashion while also offering computational efficiency and obtaining qualitatively similar results compared to traditional iterative model-based methods.