Image noise level estimation based on higher-order statistics

Abstract

Noise level estimation is a required step for many preprocessing algorithms in computer vision such as image denoising. In this paper, a model-based technique for additive white Gaussian noise level estimation is proposed via matching moments of eligible transform coefficients of a single image. We assume that noise and image signals are independent and seek to use an image transform that preserves distribution characteristics of noise. This transform should also result in coefficients with a generalized Gaussian distribution for the image itself. We use block-based discrete cosine transform (DCT) and discrete wavelet transform (DWT) which are shown to satisfy these requirements. The proposed method fits the histogram of AC coefficients of all DCT blocks or histogram of all high frequency wavelet coefficients with a generalized Gaussian distribution and attempts to estimate the noise variance through matching the estimated and true values of moments. Since the modeled distributions are symmetric and hence all odd moments are zero, our approach involves solving a nonlinear system of equations based on the method of moments using only even moments. The performance of the proposed method is compared to those of the best prevalent algorithms proposed for noise level estimation using patch-based and model-based techniques. The results on three different image databases show that the proposed scheme outperforms previous techniques in general.