Panoramic Image Quality Assessment Based on Scanpath Prediction and Bidimensional Feature Representation

Panoramic image quality assessment (PIQA) has been becoming a high-profile visual task with the public’s high attention to immersive media. Most existing PIQA metrics neglect the consumers’ viewing behavior of perceiving panoramic image quality along a scanpath in 360 degrees and the significance of both local and global feature representation, which hinders their further performance development. Inspired by these facts, we propose a no-reference PIQA metric based on scanpath prediction and bi-dimensional feature representation (SPP-BFR). The SPP module is achieved by the minimum entropy difference based fixation point estimation method proposed by us, which has the advantages of both simulating the viewing behavior and reducing the computation overhead. The bi-dimensional features are explored by a large kernel attention based local feature representation network and the self-attention mechanism based global feature representation network, respectively. Extensive experiments conducted on the publicly released datasets show the superiority of the SPP-BFR metric to the state-of-the-arts.