Robust single image noise estimation from approximate local statistics

A novel method for estimating the variance and standard deviation of the additive white Gaussian noise contained in an image will be presented. Only a single image is used to estimate the noise properties. Local image outliers are discarded, this allows us to separate the additive zero mean white Gaussian noise contained in a noisy image from the original image structure. Local variance estimates can then be calculated from the extracted noise. These local variance estimates are weak and can be influenced by misclassified image information. Robust statistics are then used to fuse the weak local variance estimates to obtain a robust global noise variance estimate. This method of estimating the noise properties is computationally efficient and provides reliable estimation results in synthetic and real-world imagery. The accuracy and processing complexity of the proposed algorithm will be compared against the current state-of-the-art noise estimators.