Adaptive feature fusion with mixture of experts for no-reference video quality assessment

This paper proposes a mixture of experts (MoE)-driven adaptive feature fusion framework (AFFVQA) for no-reference video quality assessment (NR-VQA) of in-the-wild videos. The framework is designed to enhance hierarchical feature selection and fusion through domain-specific experts and dynamic gating mechanisms. By integrating spatial-temporal experts as plug-in modules into the video backbone, our method adaptively processes multi-scale features, including low-level local textures and high-level semantic abstractions, to address diverse distortions in user-generated and AI-generated videos. The sparse gating network dynamically prioritizes domain-specific experts, optimizing feature fusion across the backbone stages. Our method is evaluated on the LSVQ, KoNVid-1k, and LIVE-VQC benchmarks, utilizing the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank Correlation Coefficient (SRCC) metrics. Extensive experiments demonstrate that AFFVQA can effectively improve the NR-VQA performance on the benchmark dataset, e.g., 1.3% gain on the KoNViD-1k dataset for FAST-VQA in terms of PLCC and SRCC, showcasing competitive performance against state-of-the-art methods.