Accurate image noise level estimation by high order polynomial local surface approximation and statistical inference

Image noise level estimation is an important step in many image processing tasks such as denoising, compression and segmentation. Although recently proposed SVD and PCA approaches have produced the most accurate estimates so far, these linear subspace-based methods still suffer from signal contamination from the clean signal content, especially in the low noise situation. In addition, the common performance evaluation procedure currently in use treats test images as noise-free. This omits the noise already in those test images and invariably incurs a bias. In this paper we make two contributions. First, we propose a new noise level estimation method using nonlinear local surface approximation. In this method, we first approximate image noise-free content in each block using a high degree polynomial. Then the block residual variances, which follow chi squared distribution, are sorted and the upper quantile of a carefully chosen size is used for estimation. Secondly, we propose a new performance evaluation procedure that is free from the influence of the noise already present in the test images. Experimental results show that it has much improved performance than typical state-of-the-art methods in terms of both estimation accuracy and stability.