Objective Hybrid Quality Assessment of Binary Images with the Use of Shallow Neural Networks

Abstract

The state-of-the-art image quality assessment methods designed for binary images are not highly correlated with subjective evaluation results, therefore one of the efficient methods to improve their performance is the application of shallow neural networks. In such an approach each elementary metric is used as the input of the network and the network is trained with subjective quality scores used as the goal function. The obtained correlation with subjective scores depends not only on the number of elementary metrics and their choice but also on the training algorithm and the network’s structure as presented in the paper.