A New Nonlocal Iterative Trilateral Filter for SAR Images Despeckling

Speckle noise, a common artifact inherent in synthetic aperture radar (SAR) imagery, significantly degrades image quality. This deterioration in quality impedes critical tasks, such as segmentation, object recognition, and other related applications. This article introduces an innovative iterative trilateral filtering approach for SAR imagery enhancement, which is distinct in its use of Haar wavelet coefficients combined with pixel-domain distance confidence levels. By incorporating nonlocal (NL) information, the proposed approach aims to augment conventional wavelet transforms, effectively utilizing imprecise pixel data through confidence-weighted aggregations. The method uniquely integrates transform-domain, spatial-domain, and statistical properties of SAR images, which significantly enhance image clarity by producing high-fidelity despeckled outputs. Compared with traditional wavelet transforms, the proposed approach focuses on the discreteness analysis of inter-patch wavelet coefficients, which allow for deeper insights and better preservation of SAR image structures. The algorithm computes single-level Haar wavelet coefficients for patch pairs and determines wavelet weights based on inter-coefficient distances. Then, the algorithm calculates guided image pixel distances and reduces patchwise Bhattacharyya distances to encode pixel distance confidence. These three weights are then consolidated for effective despeckling. To address directional blurring due to SAR image anisotropy, the conventional NL method’s square pixel vector is replaced with circular and annular vectors. The proposed approach demonstrates significant improvements in image integrity preservation, reducing speckle noise, as evidenced in our experiments with both simulated and real datasets.