Video Encoder Autotuning Using Bayesian Parameter Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video encoding systems face challenges in efficiently finding optimal encoding parameters for specific content instances due to the complexity of inter-dependent parameters and resource-intensive trial-and-error methods, leading to suboptimal results and excessive computing resource usage.
Innovation Solution
A machine learning-based approach using Bayesian optimization generates a dataset for optimal encoding parameters, trains a model to map content features to these parameters, and predicts optimal settings for new content instances, reducing the need for multiple encodings and conserving computing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If trial-and-error methods are used to find optimal encoding parameters, then encoding quality can be improved, but computing resource usage increases significantly
Solution Approach 1:
The system performs preliminary encoding with default parameters to extract features and generate a surrogate model before final encoding. This preliminary action captures content characteristics and builds a predictive model that guides parameter selection, avoiding the need for extensive trial-and-error encodings while still achieving optimal quality
Solution Approach 2:
The system creates a surrogate model that copies and approximates the complex encoding quality evaluation function. This surrogate model can quickly predict encoding quality without performing full encodings, enabling efficient parameter optimization by evaluating many parameter combinations virtually rather than through actual encodings
2Manufacturing precision
If multiple encodings are performed to find optimal parameters, then parameter optimization improves, but encoding time increases
Solution Approach 1:
The system performs a preliminary encoding pass to extract content features and build a surrogate model before the final encoding. This preliminary action enables the system to predict optimal parameters without requiring multiple full encodings, significantly reducing total encoding time while maintaining parameter optimization
Solution Approach 2:
The system uses the surrogate model to provide feedback on predicted encoding quality for different parameter combinations. This feedback mechanism guides the optimization process by indicating which parameter settings are likely to yield better results, allowing the system to converge on optimal parameters with fewer encoding iterations
3Manufacturing precision
If extensive parameter tuning is performed, then encoding quality improves, but system complexity increases
Solution Approach 1:
The system introduces a surrogate model as an intermediary between the content features and the encoding parameter selection. This surrogate model simplifies the complex relationship between numerous encoding parameters and quality outcomes by providing a tractable approximation that can be efficiently evaluated and optimized
Solution Approach 2:
The system changes the parameter space by working with content features extracted from a preliminary encoding rather than directly manipulating numerous encoding parameters. This transformation simplifies the optimization problem by reducing the dimensionality and inter-dependencies of the parameters that need to be tuned
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In some embodiments, a method determines an instance of content and a metric to evaluate a quality of an encoding of the instance of content. A set of features is extracted. The method performs an optimized search process to evaluate different combinations of encoding parameter values that are used to encode the content to generate instances of encoded content. The instances of encoded content are compared to the metric to determine a next combination of encoding parameter values to use. An optimal combination of encoding parameter values is selected that is associated with one of the instances of encoded content. Predicted encoding parameter values are output from a model using model parameters based on an input of the set of features. The method is trained using the optimal combination of encoding parameter values and the predicted encoding parameter values, wherein the model parameters are adjusted in the training.