Abstract: In response to the inadequacies of current texture-filtering methodologies, which struggle to generate high texture-structure differentiation in structural measurement imagery, and have difficulty suppressing prominent gradient textures in conjunction with maintaining structural stability, we propose a scale-adaptive texture-filtering algorithm utilizing multi-scale windows. Initially, we propose an implementation of a circular gradient operator designed to suppress robust gradient textures while simultaneously preserving intricate structures. It leverages the multi-directional, multi-structural, multi-channel, and multi-scale data encapsulated within each pixel's circular neighbourhood vicinity. This data facilitates the computation of the disparity between textures and structures, which is then imported into the frame of DIRECTIONAL ANISOTROPIC STRUCTURE MEASUREMENT (DASM), resulting in high-contrast structural measurement imagery. Subsequently, we utilize the Gaussian Mixture Model (GMM) and the Expectation-Maximization (EM) algorithm to differentiate between textures and structures. Lastly, under the guidance of multi-scale windows, we assess the likelihood of structural presence within a pixel's local vicinity. Consequently, we propose a mechanism that appraises and adjusts the scale of the filtering kernel - favoring larger kernels for textured areas and smaller ones for structural spaces, thereby ensuring scale-adaptive filtering results. In order to validate the efficacy of our proposed methodology, we facilitate a comparative analysis with existing filtering techniques. We employ F1 metric-based evaluations on traditionally-used image datasets in the image-smoothing domain and utilize PSNR and SSIM metrics for assessments on the BSD500 dataset. The inference drawn from the experiment substantiates that our algorithm is adept at suppressing robust gradient textures and preserving structural stability and aesthetics when juxtaposed with existing methods.