Online Video Quality Model for Real-Time Subjective Score Approximation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image/video quality assessment methods face challenges in achieving real-time subjective quality evaluation with high accuracy without increasing server-side hardware costs, particularly in the context of cloud gaming where massive video data requires immediate feedback.
Innovation Solution
A method involving an online model trained using offline models and subjective quality values from training samples, allowing for real-time assessment of image/video quality through an approximate value that approximates subjective truth values without increasing server hardware costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a complex quality assessment model (e.g., VAMF) is used to improve assessment accuracy, then the assessment accuracy is improved, but the calculation complexity increases significantly and real-time assessment at high frame rate and resolution is no longer achievable
Solution Approach 1:
The patent segments the quality assessment process into two distinct phases: an offline training phase where a complex model is pre-trained using historical data, and an online inference phase where a simplified model makes real-time predictions. This segmentation allows the system to achieve high accuracy during training while maintaining low complexity during real-time operation.
Solution Approach 2:
The patent applies preliminary action by pre-training a quality assessment model using historical video data and subjective quality scores before actual real-time assessment. The pre-trained model captures complex quality relationships in advance, allowing the online system to make accurate predictions without performing complex calculations in real-time.
2Measurement precision
If subjective quality scoring is performed manually to obtain accurate quality values, then the assessment accuracy is improved, but the time consumption increases and real-time assessment cannot be achieved
Solution Approach 1:
The patent creates a computational copy of human subjective quality assessment through a trained neural network model. The model learns to replicate human quality judgment patterns from training data, enabling automated prediction that matches human assessment accuracy while eliminating the time-consuming manual scoring process.
Solution Approach 2:
The patent replaces the mechanical process of manual human scoring with an automated neural network-based assessment system. The neural network processes video data and outputs quality scores automatically, substituting human cognitive evaluation with computational analysis that achieves similar accuracy much faster.
3Productivity
If real-time subjective quality assessment is implemented to provide immediate feedback, then the productivity is improved, but the hardware cost increases due to the need for powerful processing capabilities
Solution Approach 1:
The patent applies dynamics by using a simplified model architecture that adapts to real-time computational requirements. The online model is designed with reduced complexity compared to the training model, allowing it to operate efficiently on standard hardware while maintaining the ability to process video data in real-time at high frame rates.
Solution Approach 2:
The patent changes the model parameters and architecture from the complex pre-trained model to a simplified online model with fewer parameters. This parameter reduction decreases computational requirements and hardware costs while preserving the essential quality assessment functionality for real-time operation.
Data Source
AI summary
A method for assessing image/video quality includes obtaining a sample, and calculating a first approximate value approximating a subjective quality value of the sample by using an online model based on parameters of the sample. The online model is implemented by a server. The method further includes assessing quality of the video or the image of the sample based on the first approximate value. The online model is obtained based on an offline model, and the offline model is obtained by using k training samples and subjective quality values of the k training samples as a training set. The subjective quality values of the k training samples are obtained in a subjective scoring manner, and the online model is obtained by fitting parameters of the k training samples with reference to approximate values of the k training samples obtained by using the offline model.