Improved Noisy Image Quality Assessment Using Multilayer Neural Networks

Image quality assessment plays a key role to provide high Quality of Experience (QoE) in image processing applications, especially where the end user is a human. Among different distortions that may appear during image acquisition, noise is the most annoying one. In this study, we predict values of full-reference image quality metrics (FR IQA) for noisy images corrupted by additive white Gaussian noise (AWGN) without any access to the reference (noise-free) image. Prediction is carried out using energy allocation features extracted from discrete cosine transform domain as well as a priori known noise variance. Multilayer perceptrons are used to map the features from noisy images into FR IQA scores. It is shown that the use of a priori known noise variance significantly improves prediction accuracy. The influence of noise variance estimation errors on prediction accuracy is analyzed. The source codes along with the demo Android application, full datasets and supplementary materials will be available online at https://github.com/ViA-RiVaL/NoisyFRIQA-MLP.