Additive White Gaussian Noise Level Estimation for Natural Images Using Linear Scale-Space Features

摘要

Noise in images is often modelled with additive white Gaussian noise (AWGN). An accurate estimation of noise level without any prior knowledge of noisy input image leads to effective blind image denoising methods. The performance of certain image denoising methods under AWGN model is dependent on the accuracy of noise level estimation (NLE). Hence, there is a need to develop an effective NLE method in order to achieve better performance in image denoising. Even though the existing NLE methods perform well on natural images, these methods involve complex segmentation tasks such as homogeneous regions extraction and super-pixel decomposition. Hence, a simple, fast, and accurate NLE method for AWGN is proposed in this paper. In the presented NLE method, the statistical features of high-frequency details of noisy input image are obtained at multiple linear (Gaussian) scale-space which are used to construct a feature vector. It is perceived that the features obtained are almost linear and separable. Hence, supervised linear regression (LR) models that are trained globally and locally are suggested for NLE. The proposed method estimates the noise level in two stages. In stage-1, a globally trained LR model is used to estimate the noise level. It is observed that the accuracy of the noise level obtained through stage-1 can be further improved in stage-2 by adopting the proposed locally trained LR model. The proposed NLE method is evaluated with artificially generated noisy natural images using AWGN model. The high-quality natural images from Waterloo and BSD500 datasets are selected using image quality selection module and then used in training and testing phases. The average absolute deviation (AAD) is evaluated from each selected image in the datasets over a wide range of noise levels ([0 100]). The average AAD for selected images in Waterloo (BSD500) dataset is 0.21 (0.18), and execution time required to estimate the noise level is 0.04 s per image. From the obtained results, it is clear that the proposed method is simple, fast, and accurate as compared to several existing NLE methods. The effectiveness of the proposed NLE method is illustrated with fast and flexible denoising convolutional neural network using standard test images at randomly selected noise levels.