Omnidirectional Image Quality Assessment With Mutual Distillation

There exists complementary relationship between different projection formats of omnidirectional images. However, most existing omnidirectional image quality assessment (OIQA) works only operate solely on single projection format, and rarely explore the solutions on different projection formats. To this end, we propose a mutual distillation-based omnidirectional image quality assessment method, abbreviated as MD-OIQA. The MD-OIQA explores the complementary relationship between different projection formats to improve the feature representation of omnidirectional images for quality prediction. Specifically, we separately feed equirectangular projection (ERP) and cubemap projection (CMP) images into two peer student networks to capture quality-aware features of specific projection contents. Meanwhile, we propose a self-adaptive mutual distillation module (SAMDM) that deploys mutual distillation at multiple network stages to achieve the mutual learning between the two networks. The proposed SAMDM is able to capture the useful knowledge from the dynamic optimized networks to improve the effect of mutual distillation by enhancing the feature interactions through a deep cross network and generating masks to efficiently capture the complementary information from different projection contents. Finally, the features extracted from single projection content are used for quality prediction. The experiment results on three public databases demonstrate that the proposed method can efficiently improve the model representation capability and achieves superior performance.