Multi-Observer Scanpaths for Omnidirectional Image Quality Assessment

Different observers have different scanpaths when viewing an omnidirectional image. We follow this to construct scanpaths based on weighted random walk on graphs, and propose an ensemble learning framework for omnidirectional image quality assessment (OIQA) with multiple visual scanpaths of observers. First, in order to simulate the observation behaviors of different observers, we construct the Markov process by performing random walks on a weighted undirected graph. Facing the diverse visual scanpaths brought by different observers, we use the ensemble learning to treat the different observers' visual scanpaths for OIQA. The experiment results on two datasets show that the proposed method achieve better performance than compared models.