The Effect of Quantization Setting for Image Denoising Methods: An Empirical Study

Abstract

Image denoising, which aims to recover a clean image from a noisy one, is a classical yet still active topic in low level vision due to its high value in various practical applications. Existing image denoising methods generally assume the noisy image is generated by adding an additive white Gaussian noise (AWGN) to the clean image. Following this assumption, synthetic noisy images with ideal AWGN rather than real noisy images are usually used to test the performance of the denoising methods. Such synthetic noisy images, however, lack the necessary image quantification procedure which implies some pixel intensity values may be even negative or higher than the maximum of the value interval (e.g., 255), leading to a violation of the image coding. Consequently, this naturally raises the question: what is the difference between those two kinds of denoised images with and without quantization setting? In this paper, we first give an empirical study to answer this question. Experimental results demonstrate that the pixel value range of the denoised images with quantization setting tend to be narrower than that without quantization setting, as well as that of ground-truth images. In order to resolve this unwanted effect of quantization, we then propose an empirical trick for state-of-the-art weighted nuclear norm minimization (WNNM) based denoising method such that the pixel value interval of the denoised image with quantization setting accords with that of the corresponding ground-truth image. As a result, our findings can provide a deeper understanding on effect of quantization and its possible solutions.