Directional relative total variation for structure–texture decomposition

Abstract

Structure–texture decomposition is one of the fundamental branches of computer vision and image processing. It is not only beneficial to image understanding but also to subsequent object recognition and tracking. However, diversity and complexity of textures impede structure extraction. In this study, a simple yet effective structure–texture decomposition method is proposed based on directional relative total variation. More concretely, the directionality of relative total variation is taken into account by using the directional gradient operator instead of partial derivatives both in x and y directions. In addition, in order to explore the directional information on different scales, the non-subsampled pyramid filter bank is employed to obtain the multi-scale properties. Experiments on a wide variety of images reveal the efficacy of the proposed method and show its superiority over several state-of-the-art methods. Besides, the applicability of the proposed method to several image processing tasks is also demonstrated, such as edge extraction, defect detection, and so on.