Objective picture quality scale by neural networks (PQS‐NN)

Abstract

AbstractThe scale for evaluating the picture quality of coded pictures objectively becomes very important because some high‐efficiency coding techniques (e.g., DCT and VQ) are applied to the practical stage. An objective picture quality scale (PQS) has been proposed which approximates well the mean opinion score (MOS) by analyzing the perceptual properties of image distortions and using multiple regression analysis (MRA) to construct a linear combination of the essential factors of distortions extracted from the fundamental distortions by principal component analysis (PCA).In this paper, a new objective picture quality scale (PQS) is obtained by three‐layered neural networks feeding the coded error images to the input layer and giving the mean opinion scores as a target. No knowledge of visual perception is given to neural networks.After learning that the mean opinion score of open test images (out of learning) is approximated well by neural networks, it is shown that neural networks (or NN) estimate MOS well with two hidden units and that the estimate performance is improved with an increase in the number of learning images. A very interesting fact is clarified, i.e., NN learned the same functions as visual perceptions. An objective picture quality scale, with considerable precision in estimating test images without professional knowledge about perceptual properties, can be obtained by using this new objective picture quality scale PQS‐NN.