Adaptive Texture Filtering Algorithm Based on Multi-Scale Window

The current texture filtering methods often struggle to generate structural measurement image with high texture-structure differentiation and have difficulty suppressing texture with strong gradient while maintaining structural stability. To address the problem, an adaptive texture filtering algorithm based on multi-scale window is proposed. Firstly, a circular gradient operator is designed to suppress texture with strong gradient while simultaneously preserving intricate structure. It utilizes the multi-direction, multi-structure, multi-channel, and multi-scale information within each pixel’s circular neighbourhood to compute the difference between texture and structure, and then input them into the framework of Directional Anisotropic Structure Measurement (DASM) to obtain high-contrast structural measurement image. Secondly, a texture-structure separation method based on Gaussian mixture model and EM algorithm is proposed. Finally, a mechanism for adaptive filter kernel selection based on the possibility assessment of the existence of structure in the neighborhood of pixel under multi-scale window is proposed. It can enable the pixel located in the texture to obtain large-scale filter kernel, and the pixel located in the structure to obtain small-scale filter kernel, thus achieving scale-adaptive filtering effect. In the experiment, the classical images in the field of image smoothing and the images from the BSD500 dataset are used as the test data, and the F₁, PSNR, and SSIM are adopted as the objective evaluation matrics. The results show that the proposed algorithm is able to effectively suppress texture with strong gradient and preserve the stability and aesthetics of intricate structural sharp region, compared with existing texture filtering methods.