Region covariance based total variation optimization for structure-texture decomposition

With conventional structure-preserving filters, it is not always easy to remove texture details and preserve important structures from images of high complexity. To enhance the performance of structure-texture decomposition, a new method based on region covariance of simple image features and total variation optimization is proposed in this paper. We first identify texture from structure by patch-based covariance, which shows highly discriminative power for textures. Then, a total variation model built on the joint covariance and gradient is used for structure-preserving smoothing. To overcome the inherent limitation of covariance descriptor in locating main structures, patch shifting based on the variation of the region covariance is introduced. We compare our approach with state-of-the-art structure-preserving decomposition methods and the results show that our approach outperforms them in removing unimportant texture details while preserving main structures. Even for images containing the mixture of high-contrast textures with obscure boarders between them, our approach still can improve the decomposition at few extra cost of computation. Besides better decomposition results and robustness for various types of images, the simplicity of our approach make it easy to implement and adaptable to other applications.