An Improved Lee Filter Based on Physical Perception for SAR Aircraft Target Detection

Speckle noise is a fundamental challenge in aircraft target detection using synthetic aperture radar (SAR), as it degrades image quality, blurs edges, and induces false alarms. The classic Lee filter achieves a good balance between noise suppression and edge preservation, but its weights rely solely on local first- and second-order statistics, making it unable to adaptively distinguish between flat regions and complex structures. To address this, this paper proposes an improved Lee filtering method guided by local structural complexity. The coefficient of variation (CV) is introduced to quantify structural complexity, and a Sigmoid mapping is used to adaptively adjust the noise variance - enhancing smoothing in flat areas while preserving details in edge regions. Additionally, a kurtosis enhancement strategy is designed to retain weakly scattered targets. Experiments on the SAR-AIRcraft-1.0 dataset demonstrate that the proposed method improves the equivalent number of looks (ENL) from 3.481 to 3.586 and achieves an edge preservation index (EPI) of 0.978, compared to 0.973 for the classical Lee filter. As a preprocessing step for YOLOv8, the method improves mAP@0.5 by 0.6% and recall by 1.6%.