Edge Preserving Multiplicative Noise Removal of SAR Images Through Convolutional Neural Network and Anisotropic Diffusion

Synthetic Aperture Radar (SAR) is a crucial commodity due to its ability to provide high-resolution imaging across all weather conditions. Multiplicative noise often degrades the quality of SAR images. In this work, a multi-tiered approach is proposed to enhance the $S A R$ image quality. First, denoising is performed by ResNeXt through aggregated residual transformations without the extra complexity of deeper architectures. Further, a Gabor filter is applied to refine textures and edges. Anisotropic diffusion is used at last to preserve critical edges while smoothing. The proposed technique is tested on both real and synthetic datasets. The performance is evaluated using five different metrics such as Peak Signal-toNoise Ratio (PSNR), Edge Preservation Index (EPI), Equivalent Number of Looks ($E N L$), and others. It shows significant performance across all the metrics, achieving a PSNR of 50.3 dB, approximately $\gt56 \%$ than the baseline literature, with an EPI score of 0.87 dB, indicating good edge preservation.