Video Quality Assessment based Deep Meta learning

Abstract

Video Quality Assessment (VQA) is essential to ensure user satisfaction in multimedia systems. Traditional methods, whether subjective or objective, often struggle to generalize across diverse content and distortion types. While advances in deep learning have improved prediction accuracy, many models still face challenges in adapting to unseen scenarios with limited labeled data. This thesis proposes an innovative no-reference video quality assessment approach using the Model-Agnostic Meta-Learning (MAML) algorithm, referred to as VQA-DML (Video Quality Assessment based on Deep Meta-Learning). The objective is to train a regression model capable of predicting subjective quality scores (MOS) by learning from multiple small-scale tasks drawn from various video databases. Experiments were conducted on three well-known datasets: LIVE, KoNViD-1k, and YouTube-UGC, covering a wide range of distortions and content types. The model's performance was evaluated using standard regression metrics (MAE, RMSE, R²) and correlation coefficients (PLCC, SRCC, KRCC), highlighting its ability to generalize effectively to new videos. This work offers a flexible and adaptive framework for video quality prediction, with promising prospects for real-world multimedia applications. The proposed VQA-DML model achieved promising results, with Pearson correlation up to 0.6548 and Spearman correlation up to 0.6679, outperforming several traditional NR-VQA models. These results demonstrate the model’s ability to generalize effectively to unseen video content with minimal