Quantizing Neural Networks with Knowledge Distillation for Efficient Video Quality Assessment
Jia Yuan Yu, Yingming Li · 2024
No-Reference Video Quality Assessment (NR-VQA) refers to evaluating the perceptual quality of videos without access to the original reference videos. However, existing deep NR-VQA methods often entail significant computational overhead, posing challenges in deploying these models in practical applications. In this paper, we focus on developing lightweight no-reference video quality assessment models and propose an Enhanced Quantization with Knowledge Distillation (EQKD) strategy to perform model compression while maintaining competitive performance. Specifically, we first introduce quantization-aware training in the VQA model parameter learning, which significantly reduces the memory footprint of the VQA models. Additionally, we incorporate knowledge distillation into the training process to further enhance the performance. The experimental results on three public datasets demonstrate the superiority of the proposed models over the state-of-the-art.