Accuracy of Model-based and Learning-based Approaches for Image Noise Variance Estimation
Mikhail Uss, Владимир Васильевич Лукин, Benoît Vozel, Kacem Chehdi · 2020 IEEE Ukrainian Microwave Week (UkrMW) · 2020
This paper deals with the problem of noise parameters' estimation from noisy image patches. We consider two estimators - model- and learning-based ones - with ability to estimate noise standard deviation (SD) and predict confidence of this estimate individually for each image patch. The model-based approach represents maximum likelihood estimator (MLE) of fractional Brownian motion (fBm) field parameters, and learning-based is NoiseNet convolutional neural network (CNN) trained on real-life images. We compare efficiency and ability to predict noise SD estimates confidence in two domains: pure fBm data and real-life images. We demonstrate that the learning-based approach is less effective on pure fBm data due to bias, and the model-based approach fails for complex image patches from real-life images. Based on analysis of this paper, usage of synthetic fBm data could be suggested as additional source of training data for learning-based methods of noise parameters estimation.