An Adaptive Wavelet Image Denoising Scheme Based on Local Variance

Local variance is often used to separate homogeneous and textured regions in natural image denoising. In this paper, we present an adaptive wavelet denoising scheme based on the distinction of local oscillatory property between textures (or small details) and noise at different scales. In the scheme, the local variance information is used to partition the image domain into “noisy smooth” dyadic regions and “texture or edge + noise” dyadic regions. In “noisy smooth” regions, noise can be easily removed by shrinking the corresponding high frequency wavelet coefficients to zero. In texture regions, the noise can be suppressed by multiplying a damping factor, which is relative to the local variance of the region and therefore is space adaptive. We also give the detailed algorithm to select the optimal threshold parameter in the sense of PSNR and realize multi-scale denoising. Finally, we display some numerical experiments to demonstrate the potential of our method.