Graph-based No-Reference Video Quality Assessment Using Spatial Features

In this work, we explore the efficacy of graphs for representing video sequences in the context of no-reference video quality assessment (NR-VQA). Specifically, we address the problem of predicting video quality from frame-level spatial quality scores by using them as node features. We implicitly capture temporal dependency among the nodes through an appropriate choice of the graph adjacency matrix. A graph convolutional network (GCN) that accepts the graph video representation as its input is trained to predict mean opinion scores (MOS). We experiment with several NR image quality assessment (NR-IQA) algorithms that include Natural Image Quality Evalua-tor (NIQE), Multi-scale Image Quality Transformer (MUSIQ), Codebook Representation for No-Reference Image Assessment (CORNIA), and Neural Image Assessment (NIMA) for predicting spatial quality. The proposed NR-VQA algorithm is evaluated on four user-generated VQA datasets on which it showed a consistent improvement over directly using spatial quality scores for regression. These improvements suggest that graph-based video representation is promising for the NR-VQA task.