Investigating SAR Data Denoising: A Comparative Analysis of CNN Models with Multi-Channel Signal Processing Features

This paper investigates the necessity for denoising in Synthetic Aperture Radar (SAR) dataset, emphasizing the comparison of three distinct neural networks: U-Net, ResNet, and DeepDeblur. Utilizing the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset, augmented with additive colored noise, and subjected to transformations such as rotation and shear along both axes, the proposed work applies signal processing techniques to generate additional features. These include Fast Fourier Transform (FFT) real and imaginary components, as well as Power Spectral Density (PSD), which serve as multiple-channel inputs across various experimental trials. These trials encompass scenarios ranging from using solely the noisy image to incorporating combinations involving FFT components and PSD. To assess network performance, a variety of regression metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Coefficient of Determination (COD), and Root Mean Square Error (RMSE) are employed. Subsequently, the most effective denoising network is evaluated on an unseen test dataset. This dataset is augmented with elementwise multiplicative noise, simulating the characteristic speckle commonly observed in SAR datasets. The results highlight the remarkable performance of the Convolutional Neural Network (CNN) model in effectively denoising images corrupted with extrinsic multiplicative noise, along with inherent MSTAR speckle. This demonstrates its robustness and efficacy in real-world scenarios, particularly in handling intentionally introduced noise variations.