Neural Network Generator for Video Encoding Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video coding technologies, such as HEVC, face challenges in efficiently encoding and decoding images due to the complexity of intra and inter prediction processes, which increases processing burden and configuration complexity.
Innovation Solution
The use of a neural network-based generator network in both encoders and decoders to generate predicted images and calculate prediction errors, simplifying the prediction process and improving encoding efficiency through generative adversarial networks (GANs) for intra and inter prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If traditional intra and inter prediction processes are used in HEVC, then encoding accuracy is maintained, but processing burden and configuration complexity increase
Solution Approach 1:
The patent replaces the traditional mechanical prediction process (intra prediction using reference blocks and inter prediction using motion compensation) with a neural network-based generator network. The generator network automatically learns prediction patterns from training data, substituting the complex mechanical algorithms with an intelligent system that produces predicted images more efficiently while maintaining or improving accuracy.
Solution Approach 2:
The patent changes the fundamental parameters of the prediction process by transitioning from deterministic algorithmic operations to probabilistic neural network inference. The generator network uses learned weights and biases to generate predictions, fundamentally altering how prediction is performed from rule-based to model-based, thereby reducing processing burden while maintaining accuracy.
2Manufacturing precision
If complex prediction algorithms are used to improve encoding accuracy, then manufacturing precision improves, but device complexity increases
Solution Approach 1:
The patent merges the encoder and decoder prediction processes by using the same generator network in both devices. This unified approach ensures that both encoder and decoder produce identical predicted images, simplifying the overall system configuration while maintaining high encoding accuracy. The merging eliminates the need for separate, complex prediction algorithms in each device.
Solution Approach 2:
The patent uses copying by deploying the trained generator network model to both encoder and decoder devices. The same network architecture, weights, and parameters are copied to both ends of the communication channel, ensuring consistent prediction behavior and reducing configuration complexity while achieving high encoding accuracy.
3Productivity
If traditional prediction methods are used, then configuration simplicity is maintained, but encoding efficiency decreases
Solution Approach 1:
The patent applies preliminary action by training the generator network offline before actual encoding operations. The network is pre-trained on large datasets to learn optimal prediction patterns, and this pre-trained model is then deployed for efficient real-time encoding. This preliminary training phase separates the complex learning process from the operational phase, improving encoding efficiency while managing configuration complexity.
Data Source
AI summary
The encoder includes processing circuitry, and memory. Using the memory, the processing circuitry: generates a predicted image of an input image that is a current image to be encoded, based on generated data output from a generator network in response to a reference image being input to the generator network, the generator network being a neural network; calculates a prediction error by subtracting the predicted image from the input image; and generates an encoded image by at least transforming the prediction error.