Multi-Neural Network Models for MOS Prediction in Mobile Video Streaming

Mobile video streaming quality assessment is crucial for ensuring an optimal user experience. Traditional objective metrics fail to capture subjective video quality effectively. This article presents a method for predicting video Mean Opinion Score (vMOS) using five types of neural networks: Narrow, Medium, Wide, Bilayered, and Trilayered. The models were evaluated based on prediction accuracy, inference speed, training time, and model size. Results show that the Narrow model achieves the highest accuracy on the test set, with a Root Mean Squared Error (RMSE) of 0.0266 and a Coefficient of Determination (R²) of 0.9995. In contrast, the Wide model delivers the highest prediction speed, but at the cost of reduced accuracy, longer training time, and the largest model size. The Bilayered model offers a balanced trade-off between speed, complexity, and accuracy, while deeper architectures such as the Trilayered model provide limited performance gains despite higher computational cost. Overall, the findings reveal that increasing model width enhances prediction speed but does not improve accuracy, whereas deeper models increase training time without significant benefits. These results highlight the importance of selecting model architecture based on application-specific requirements, with the Narrow model suited for accuracy-critical tasks and the Wide or Bilayered models preferred for high-throughput scenarios.