An Improved Wavelet-Based Approach for Estimating the Variance of Noise in Images

Abstract In the paper an improved approach based on Donoho’s method is proposed to estimate the variance of Gaussian noise in images. The proposed approach exploits the interscale correlation in wavelet domain to estimate the coefficients of image signals, and then get the purer noise coefficients by deducting the estimated values from the coefficients of the noisy image, then does estimation once again using Donoho’s formula with the updated data. Simulation results are presented which indicate that the proposed approach outperforms other methods, especially in scenarios where there is less noise or much more image details.