HSCIF: Hybrid Surface-Constrained Image Filtering
Fumin Wu, Caofei Luo, Hao Xu, Xinwei Zhang, Ryan Wen Liu · 2024
Image filtering plays a fundamental role in many disparate computer vision and graphics applications, which attempts to retain the salient structures (e.g., edges and contours) and remove the fine details and textures at different scales. Motivated by the success of L0-norm on gradient sparsity promotion, L0-regularized filtering methods have produced impressive imaging results. The filtered images, only using L0-norm gradient minimization, easily suffer from various artifacts, e.g., over-sharpening effects and under-filtering of high-amplitude textures, etc. To eliminate these limitations, we propose a hybrid geometry-inspired regularizer contributing to a more powerful L0-regularized filtering method. This regularizer is presented by combing the hybrid surface area constraints in both image and gradient domains. It is more robust to remove different unwanted artifacts, especially in textured and homogeneous regions. The proposed hybrid surface-constrained filtering model is thus capable of preserving salient structures while removing fine details and textures. An alternating direction method is then suggested to solve the resulting nonsmooth nonconvex optimization problem effectively. We demonstrate that our method outperforms several state-of-the-art methods in image filtering. Several practical applications of image filtering have also been presented to illustrate the effectiveness of the proposed method.