RankDVQA-mini: Knowledge Distillation-Driven Deep Video Quality Assessment

Deep learning-based video quality assessment (deep VQA) has demonstratedsignificant potential in surpassing conventional metrics, with promisingimprovements in terms of correlation with human perception. However, thepractical deployment of such deep VQA models is often limited due to their highcomputational complexity and large memory requirements. To address this issue,we aim to significantly reduce the model size and runtime of one of thestate-of-the-art deep VQA methods, RankDVQA, by employing a two-phase workflowthat integrates pruning-driven model compression with multi-level knowledgedistillation. The resulting lightweight full reference quality metric,RankDVQA-mini, requires less than 10full version (14prediction performance that is superior to most existing deep VQA methods. Thesource code of the RankDVQA-mini has been released athttps://chenfeng-bristol.github.io/RankDVQA-mini/ for public evaluation.