Video quality prediction and classification using XGBoost under variable encoding and network conditions

This study presents a machine learning-based framework for video quality assessment (VQA). This framework enables the mapping of objective metric outputs, such as Structural Similarity Index Measure (SSIM) and Video Multimethod Assessment Fusion (VMAF), to subjective Mean Opinion Score (MOS). The absence of a standardized conversion from objective VQA methods to MOS poses challenges for consistency and comparability across different evaluation methods. The proposed mapping function delivers real-time estimation of user-perceived video quality, offering an effective alternative to traditional subjective tests, which are time-consuming, costly, and unsuitable for online services such as IPTV or streaming platforms. All constructed models were trained using the XGBoost algorithm. The dataset comprises over 700 distorted video sequences, spanning the two most prevalent codecs (H.264 and H.265), along with a wide range of resolutions, bitrates, and packet loss rates. Although real-world network traffic is affected by various impairments, including delay and jitter, packet loss is the primary factor driving video quality degradation. The SSIM-to-MOS and VMAF-to-MOS regression models achieved a Pearson correlation of 0.95 and an RMSE of 0.31. Additionally, we propose a reduced-reference classifier that estimates video quality using a limited set of video characteristics, including a frame (packet) loss rate derived from the original sequence. This approach enables the model to evaluate video quality without requiring a direct comparison between the original and the test sequence. The model achieves a weighted F1 score of 0.92 and is suitable for deployment in time-sensitive services such as live streaming and IPTV, even in environments where packet loss may occur. Benchmarking against a Back Propagation Neural Network (BPNN) and state-of-the-art methods confirmed the superiority of the described approach. All source codes are publicly available to support reproducibility and further research in Quality of Experience (QoE) assessment.