Adaptive dynamic guided image filtering for edge preservation and structure extraction

Abstract

Image smoothing is a fundamental technique in the field of image processing. Different tasks require different smoothing characteristics. However, the smoothing behavior of a given operator is typically fixed and often cannot simultaneously satisfy multiple tasks, such as detail enhancement, clip-art compression artifacts removal, and structure–texture decomposition. To address this limitation, we propose an adaptive dynamic guided image filtering model that preserves edges and extracts structures. Specifically, a weight-guided map is first constructed using a local filter and then employed to iteratively guide a global optimization model. An adaptive penalty function is introduced to enhance flexibility, enabling the model to address diverse smoothing tasks by tuning the parameters accordingly. This allows it to tackle more challenging problems, such as precise structure–texture separation, which previous methods struggle with. Furthermore, we provide an efficient numerical solution to the proposed model and analyze the convergence of the iterative algorithm through experiments, demonstrating stable convergence under various parameter settings. To quantitatively evaluate the performance, we construct an Image Smoothing (IMS) dataset. Extensive experiments across various applications validate the effectiveness and superiority of the proposed algorithm.