Accurate Image Noise Level Estimation through DCT Transformation and Approximation by Chi-Square Distribution

Accurate noise level estimation is a very important step for many image denoising and computer vision tasks. So far the key idea of most methods is to separate the noise-free image part from noisy images and then use the noise part to estimate noise level. But the noise part usually still contains some image content, which often causes biased estimation results, especially in the low noise situation. In this paper, we propose a novel noise level estimation method. This method first uses the DCT basis to approximate the noise-free image part, then uses the ?2 distribution to approximate the local noise variance. Contrary to the conventional idea that the noise residual approximation can not produce accurate estimates, we show that because of the good approximation capability of the DCT basis for natural image content and the ?2 distribution approximation used in the statistical inference, it is possible to achieve more accurate image noise estimation results than most sophisticated state-of-the-art methods especially at low noise level which is more meaningful. Experiments results confirm the advantages of the proposed approach.