Presented Quality Metrics for Media Streaming

Instead of measuring quality as distortion from reference (VMAF approach), the patent inverts the approach by measuring quality as absolute perceptual quality using trained models. The system trains deep neural networks to predict perceptual quality scores directly, rather than calculating distortion metrics. This inversion resolves the contradiction by making the measurement reflect actual human perception rather than mathematical deviation from source.

The patent changes the measurement parameter from distortion-based metrics to perceptual quality scores generated by trained neural networks. The system uses multiple trained models (e.g., for video, audio, and combined quality) that output perceptual quality scores on a standardized scale, fundamentally changing how quality is quantified from relative distortion to absolute perceptual assessment.

The patent applies partial action by selectively applying high bitrate only where perceptually necessary. The system evaluates each segment's perceptual quality score and bandwidth consumption, then makes intelligent decisions about which segments require higher quality encoding. This resolves the contradiction by avoiding excessive bitrate allocation across all content while maintaining quality consistency where it matters most.

The system changes from fixed bitrate allocation to dynamic bitrate adjustment based on perceptual quality scores. By using trained models to assess actual perceptual quality, the system can adjust bitrate parameters dynamically to maintain consistent perceived quality while optimizing bandwidth usage according to actual content characteristics rather than uniform allocation.

The patent segments the quality assessment system into specialized trained models for different content types (video quality model, audio quality model, combined quality model). Each model is trained independently on specific data types and can be applied selectively based on content format. This segmentation reduces overall system complexity by allowing modular deployment rather than requiring a single complex universal model.

The system uses trained neural network models that can be copied and deployed across multiple devices and services. Once a quality model is trained on reference data, the trained model weights and architecture can be replicated and used independently, reducing the complexity burden on individual devices while maintaining high prediction accuracy through centralized model training and distribution.