Multi-Metric Fusion Network for Image Quality Assessment

Abstract

With the fast proliferation of multimedia applications, the reliable prediction of image/video quality is urgently needed. Many quality assessment metrics have been proposed in the past decades with various complexity and consistency with human ratings. The metrics are designed from different aspects, e.g., pixel level fidelity, structural similarity, information theory and data-driven. In this paper, we design a Multi-Metric Fusion Network (MMFN) for aggregating the quality scores predicted by diverse metrics to generate more accurate results. To be specific, we utilize the image features extracted from the pretrained network to adaptively rescale the predicted quality from different metrics, and leverage the fully-connected layers to regress a single scalar as the final score. Pairwise images can be further integrated into the training procedure by adding a Score2Prob layer. Experimental results on the validation and test sets demonstrate that our proposed MMFN achieves better prediction accuracy compared with other metrics.