Improved NL-means algorithm based on gradient grouping for SAR image despeckling

Recently, a non-local means (NL-means) algorithm is proposed for the denoising of Gaussian noises, and it can effectively preserve textures while removing noises. However, the computation of NL-means is extremely heavy, and furthermore, the Gaussianity of the noise is required. Due to these limitations, NL-means is not suitable for synthetic aperture radar (SAR) image despeckling. A similar points matching method based upon gradient grouping is presented, and this method leads to an improved NL-means model. Compared with the existing NL-means, the improved model can achieve a better despeckling performance with lower computation. On the other hand, since SAR image is with multiplicative noise, the holomorphic transform is introduced as a pre-processing to cater for SAR image despcekling. Experimental results demonstrate that the proposed method achieves 3 dB of PSNR gain in comparison with the existing relevant methods and the computation is around 3 times faster than that of the NL-means.