Opinion-unaware blind quality assessment of AI-generated omnidirectional images based on deep feature statistics

Abstract

The advancement of artificial intelligence generated content (AIGC) and virtual reality (VR) technologies have prompted AI-generated omnidirectional images (AGOI) to gradually into people’s daily lives. Compared to natural omnidirectional images, AGOIs exhibit traditional low-level technical distortions and high-level semantic distortions, which can severely affect the immersive experience for users in practical applications. Consequently, there is an urgent need for thorough research and precise evaluation of AGOI quality. In this paper, we propose a novel opinion-unaware (OU) blind quality assessment approach for AGOIs based on deep feature statistics. Specifically, we first transform the AGOIs in equirectangular projection (ERP) format into a set of six cubemap projection (CMP)-converted viewport images, and extract viewport-wise multi-layer deep features from the pre-trained neural network backbone. Based on the deep representations, the multivariate Gaussian (MVG) models are subsequently fitted. The individual quality score for each CMP-converted image is calculated by comparing it against the corresponding fitted pristine MVG model. The final quality score for a testing AGOI is then computed by aggregating these individual quality scores. We conduct comprehensive experiments using the existing AGOIQA database and the experimental results show that the proposed OU-BAGOIQA model outperforms current state-of-the-art OU blind image quality assessment models. The ablation study has also been conducted to validate the effectiveness of our method.