PROJECT TITLE :

Structure-Texture Image Decomposition Using Deep Variational Priors

ABSTRACT:

Structure-texture image decomposition invariably requires structure pictures to have modest norms in some functional spaces and to share a common idea of edges, i.e., large gradients or intensity differences. As a result of this definition, it's difficult to discern between structure edges and oscillations that have fine spatial scale but strong contrast. A new model is introduced in this research that learns deep variational priors for structure images without explicit training data. As an iterative smoothing process is implemented, the modular structure of a multiplier algorithm and its alternate direction approach are used to plug deep variational priors into the process. CNNs can be used to replace the total variation prior and are powerful enough to capture the natures of structure and texture. For example, we show that our CNN trained pre-conditions can correctly identify a high-amplitude feature from a structural edge. Although prior data-driven smoothing schemes were limited in their ability to provide continuous smoothing effects, our formulation gives an additional level of flexibility. Various computational photography and Image Processing applications, such as texture removal, detail manipulation, HDR tone mapping, and non-photorealistic abstraction, have been shown to benefit from our technique.