Blind Omnidirectional Image Quality Assessment With Viewport Oriented Graph Convolutional Networks

Quality assessment of omnidirectional images has become increasingly urgent due to the rapid growth virtual reality applications. Different from traditional 2D and videos, contents can provide consumers with freely changeable viewports a larger field view covering 360 ° ×180 spherical surface, which makes objective quality more challenging. In this paper, motivated by characteristics human vision system (HVS) viewing process contents, we propose novel Viewport oriented Graph Convolution Network (VGCN) for blind image (IQA). Generally, observers tend give subjective rating 360-degree after passing aggregating different information when browsing scenery. Therefore, in order model mutual dependency image, build spatial viewport graph. Specifically, graph nodes are first defined selected higher probabilities be seen, is inspired HVS that beings sensitive structural information. Then, these connected relations capture interactions among them. Finally, reasoning on proposed performed via convolutional networks. Moreover, simultaneously obtain global using entire without sampling boost performance according experience. Experimental results demonstrate our outperforms state-of-the-art full-reference no-reference IQA metrics two public databases.