On Prediction of Image Denoising Expedience Using Neural Networks

Assessment of image visual quality is important for different stages of image processing chain. One of them is denoising that can either improve image quality or remain it the same (or even degrade). Our paper concerns the question on whether or not the use of denoising is expedient for a given image. A common model of additive white Gaussian noise (AWGN) is used. Analysis of denoising expedience is based on prediction of visual quality for acquired noisy image. A peculiarity of the proposed approach is that it employs a trained neural networks (NN) and does not require a priori known noise variance. The NN uses features extracted from spectral domain as inputs and carries out prediction of visual quality metrics for noisy image. The prediction is much faster than denoising itself and appropriately accurate. Besides, this paper evaluates the existing full-reference image quality metrics with application to image denoising using Spearman rank-order correlation coefficient (SROCC). The best SROCC values do not exceed 0.81. The image database and codes are available online at https://github.com/ViA-RiVaL/Image-Denoising-Expedience/.