Abstract: Aiming at the problem that image texture filtering need to preserve structures while filtering multi-scale texture details, we proposed a  weighted least squares texture filtering based on structure tensor. The method used anisotropy of structure tensors and local gradient to extract the structure information in the image, so as  to optimize the smoothing weights of each pixel in the 
traditional weighted least squares filtering, which determined the smoothness degree of each pixel. In the global optimization framework of minimization, the constraints of the pixels at the structures were relatively small, so smaller weights were assigned,  the constraints of the pixels in the texture regions were relatively large, so larger weights were assigned. The constructed multi-scale edge-preserving smoothing algorithm framework could decompose the image into a segmented smooth base layer and texture detail layer to obtain the edge-preserving smoothing result of natural image, so as to realize the purpose of structure-aware adaptive texture filtering. The experimental results of  subjective and objective comparison with the common methods show that the proposed algorithm can effectively maintain the structure edges while smoothing out multi-scale textures, and has obvious advantages in terms of structure similarity and visual perception quality.

Key words: texture filtering, structure-preserving, structure tensor, weighted least squares filtering