Numba-based Non-local Means Filter for SAR Image Denoising

Synthetic Aperture Radar (SAR) images are an essential tool for earth observation and remote sensing, but speckle can degrade the quality of such images. Image processing techniques such as filtering and averaging can effectively reduce the impact of noise on images. This paper evaluates the performance of the hybrid nonlocal means (NLM) filter in denoising Sentinel-1 SAR images. The proposed approach combines adaptive exponential kernel-based pre-processing with an NLM filter. The NLM algorithm is computationally intensive and time-consuming due to its pixel-by-pixel denoising based on extended neighborhood similarities. To address this challenge, this paper explores parallelization of the NLM algorithm using NUMBA, a Python compiler, to accelerate its performance. It is found that the proposed approach shows a significant improvement in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Mean Squared Error (MSE) and performs better than conventional NLM filters and even provides the greatest speedup when compared to earlier parallel NLM techniques.