Neural Network-based Image Quality Comparator without Collecting the Human Score for Training

Emulating human behaviours in automated image quality assessment (IQA) enables a comparator framework to remove the differences in human bias naturally. Based on the observation of the practical applications of IQA, this study focuses on similar-content image quality comparison based on a new image quality comparator (IQC). Outstanding proven IQAs can be utilised in this comparator to achieve a new non-linear combination strategy to boost the IQAs' performance in image quality comparison. For both input images to be compared, proven IQAs are utilised to obtain nine features from each image, yielding 18 total features. Then, a four-layer comparison network conducts a classification task to indicate which input image has better quality. In the training phase, the commonly used human scores as training labels are replaced with pairwise comparison results that are automatically generated from assigned distortion level differences. By not utilising human score in training phase, this IQC shows two advantages: (i) it removes huge labor and time cost to collect the human scores and (ii) it solves the problem of over-fitting benefiting from simplicity of creating a large image training dataset. Furthermore, the experimental tests and cross-dataset validation comparison tests demonstrate its impressive performance.