Neural Network Decoder Training With Quantization for Codec Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural-based video codecs lack sufficient compression efficiency, and transmitting dedicated decoders with encoded signals can cancel the gains achieved during training, leading to high bitrate and reduced decoding performance.
Innovation Solution
A method and apparatus for training a neural network-based decoder using quantization during the training process, allowing the decoder to learn correct decoding in the presence of quantization noise, enabling more efficient compression without requiring additional parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the decoder is trained without quantization, then decoding performance is maintained, but compression efficiency is reduced
Solution Approach 1:
The patent applies quantization during the training phase before actual deployment. By pre-training the decoder with quantized parameters, the system prepares the model to handle quantization effects, enabling efficient compression while maintaining performance. This preliminary action resolves the contradiction by anticipating the quantization that will occur during transmission.
Solution Approach 2:
The patent transforms the harmful effect of quantization noise into a beneficial training condition. By exposing the decoder to quantization during training, the model learns to robustly handle quantized inputs, converting what would normally be a performance-degrading factor into a training advantage that improves compression efficiency.
2Measurement precision
If additional parameters are used to maintain decoding performance, then decoding accuracy is improved, but device complexity increases
Solution Approach 1:
The patent changes the state of existing parameters by applying quantization during training, rather than introducing new parameters. This approach maintains decoding accuracy while avoiding increased device complexity, as the same parameter set is used but trained under quantization conditions that improve its efficiency.
3Loss of substance
If the decoder is compressed more, then compression efficiency is improved, but decoding performance deteriorates
Solution Approach 1:
By pre-training with quantized parameters, the decoder is prepared for the compression that will follow. This preliminary exposure to quantization effects enables the model to maintain performance even when heavily compressed, resolving the trade-off between compression efficiency and decoding performance.
Solution Approach 2:
The training process with quantization acts as a cushion against the performance degradation that would normally result from compression. By anticipating and preparing for quantization effects during training, the system creates a buffer that protects decoding performance even when the decoder is heavily compressed for efficient transmission.
Data Source
AI summary
A device and a method for training a neural network based decoder. The method includes during the training, quantizing, using a training quantizer, parameters representative of the coefficients of the neural network based decoder. A method and device are also provided for encoding at least parameters representative of the coefficients of a neural network based decoder. Provided also are a method for generating an encoded bitstream including an encoded neural network based decoder, a neural network based encoder and decoder, and a signal encoded using the neural network based encoder.