MO-QoE: Video QoE using multi-feature fusion based Optimized Learning Models

摘要

The escalating demand for video content and streaming services has made it a predominant medium of exchanging information in the modern era. Videos are processed, compressed, and streamed over dynamic wireless channels having limited bandwidth. This introduces video impairments that deteriorate the video quality. Humans are the targeted observers of multimedia content. Improving viewing experience requires human intervened subjective assessment of video quality; which is relatively cumbersome and resource intensive. Devising alternatives that can estimate subjective quality scores are essential to evaluate Quality-of-Experience (QoE) in multimedia applications and services. The features (objective quality metrics and impairment factors) provide a measure of artifacts and content characteristics that influence observers viewing experience. We have proposed MO-QoE, a framework to predict video QoE using Multi-Feature Fusion (MFF) based Optimized Learning Models (OLMs). The OLMs are the optimized neural network models designed for QoE prediction. We make use of Adaptive Moment estimation and Batch Gradient Descent algorithms that update the values of weight and biases of the learning models to their optimal values. The inputs to our model is based on MFF scheme, that eliminates distortion specific biased performance of any individual feature. The resultant non-linear regression score from the model is obtained after fusion of scores from multiple features. The QoE prediction using the MFF-based OLMs has been tested on publicly available databases. The prediction performance/accuracy is improved by ≈2 to 3 times as compared to the existing models.