Texture filtering tries to eliminate unimportant fine-scale details for well presenting the prominent contents in images. This is an essential operation with a wide range of applications in computer vision and graphics, including scene understanding, visual abstraction, edge extraction, etc. However, unlike existing filtering techniques for removing spot noises and geometric fine details, texture filtering faces a big challenge as it is not easy to detect the texture regions. This is because a texture is always related to a visual pattern, and the pattern may be of different sizes and shapes, causing texture measurement very difficult.

Recently, a group led by Prof. WANG Wencheng of State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, presented a novel texture filtering method to smooth out multi-scale textures while preserving image structures. In the method, Dr. XU Panpan, a post doctor, and Wencheng Wang suggested to initialize large box windows for smoothing out large-scale textures, while adaptively shrinking the filtering boxes by the distances from their covered pixels to the structure edges. This is based on the observation that the pixel nearer the edges always has fewer similar pixels in its neighborhood. Thus, adaptive shrinking can be performed by checking the similar pixels to the pixel under investigation. In this way, the pixels near structures can be handled in smaller boxes, helpful for preserving structure edges and effectively removing textures. As illustrated in the following images, the proposed method can considerably improve the filtering results. The method is simple to implement, and thus it is also very efficient.

Figure : Comparison of texture filtering results with the state-of-the-arts. SSBF (J. Jeon, H. Lee, H. Kang, and S. Lee, “Scale-aware structure-preserving texture filtering,” Comput. Graph. Forum (Proc. Pacific Graphics’2016), vol. 35, no. 7, pp. 77–86, 2016.). TIF (P. Xu and W. Wang, “Improved bilateral texture filtering with edge-aware measurement,” IEEE Trans. Image Process., vol. 27, no. 7, pp. 3621–3630, 2018.)

The work entitled “Structure-aware Widow Optimization for Texture Filtering” was published in IEEE Transactions on Image Processing, Vol.28, No.9, 2019. (https://ieeexplore.ieee.org/document/8666743) This study was supported by the National Natural Science Foundation of China.