Video Coding Filter Training With Iterative Model Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding and compression algorithms rely on manually designed filtering coefficients, leading to inaccurate reduction of image distortion and poor image quality in compressed videos, especially at high compression rates.
Innovation Solution
Iteratively update a filtering model by generating training data through a video coding application, training a second filtering model based on sample original and reconstructed frames, and integrating it into the application until a filtering quality requirement is met, enhancing filtering performance without manual intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If manually designed filtering coefficients are used in loop filter, then the filtering process is simple to implement, but the image distortion cannot be reduced effectively and image quality deteriorates
Solution Approach 1:
The filtering model automatically generates its own training data by utilizing reconstructed frames from the video coding process, eliminating the need for external manual data preparation. The system trains itself iteratively by feeding its own output back as training input, achieving self-improvement without external intervention.
Solution Approach 2:
The patent implements a feedback mechanism where the reconstructed frames (which contain distortion information) are fed back as training data to update the filtering model. This closed-loop feedback allows the model to learn from its own performance and continuously improve distortion reduction capability.
2Measurement precision
If manually designed filtering coefficients are used, then the implementation process is straightforward, but the accuracy of distortion reduction is insufficient
Solution Approach 1:
The patent replaces manual mechanical design of filtering coefficients with an automated machine learning-based filtering model. The system uses neural network or deep learning approaches to automatically learn optimal filtering parameters from training data, substituting human expert manual tuning with automated computational methods.
3Manufacturing precision
If iterative updates of video coding application are performed, then filtering performance improves, but the development and deployment complexity increases
Solution Approach 1:
The patent merges the training data generation process with the video coding application itself. By utilizing reconstructed frames directly from the coding process as training data, the system combines the coding and training functions into a unified workflow, reducing the need for separate external training systems.
Data Source
AI summary
A processing method includes generating first training data through a kth updated video coding application, the first training data including a sample original video frame as a training label, and a first sample to-be-filtered reconstructed frame output from the kth updated video coding application and corresponding to the sample original video frame. The method further includes, based on the sample original video frame and the first sample to-be-filtered reconstructed frame, training a second filtering model in the kth updated video coding application to obtain the second filtering model in a training convergence state, and integrating the second filtering model in the training convergence state into the kth updated video coding application to obtain a (k+1)th updated video coding application. The method further includes, when the (k+1)th updated video coding application meets a filtering quality requirement condition, determining the (k+1)th updated video coding application as a target video coding application.