Generative Adversarial Compression for Variable Bit Rate Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression methods face challenges in achieving variable bit rate compression without the need for multiple model trainings, leading to high complexity, time, and parameter costs, while also maintaining subjective quality and semantic information preservation.
Innovation Solution
A variable bit rate generative compression system using a coding-decoding full convolution network with generative adversarial training, where the variance of feature maps is quantized based on channel redundancy, allowing a single model to perform compression across different bit rates, and entropy rate is estimated using the L1 norm of the variance, enabling adaptive hyper-parameter adjustment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple models are trained for different compression rates, then variable bit rate compression is achieved, but training complexity and time costs increase significantly
Solution Approach 1:
The patent applies universality by designing a single compression model that can operate at multiple bit rates. The model uses a shared encoder-decoder architecture with learnable quantization parameters that can be dynamically adjusted during inference to achieve different compression rates without requiring separate trained models for each rate.
Solution Approach 2:
The patent implements dynamics by introducing dynamic quantization parameters that are learned during training and can be adaptively adjusted during inference. The quantization step size and other compression parameters are not fixed but can be dynamically modified based on the desired bit rate, allowing the single model to adapt to different compression requirements.
2Productivity
If quantization method is changed to achieve variable bit rate, then compression rate is adjusted, but performance loss becomes unpredictable
Solution Approach 1:
The patent applies parameter changes by modifying quantization parameters (such as quantization step size) in a controlled and learnable manner. Instead of changing the fundamental quantization method, the approach adjusts specific parameters within the quantization process that have been optimized during training, ensuring that performance remains stable while achieving variable compression rates.
Solution Approach 2:
The patent implements feedback mechanisms where the model monitors reconstruction quality and bit rate during inference, and dynamically adjusts quantization parameters accordingly. This closed-loop approach ensures that performance degradation is minimized while achieving the target compression rate, making the process reliable and predictable.
3Manufacturing precision
If manual quantization adjustment is performed for each compression rate, then compression performance is optimized, but operational complexity and time increase
Solution Approach 1:
The patent applies self-service by enabling the compression model to automatically adjust its quantization parameters during inference based on the desired bit rate. The model performs self-optimization without requiring manual intervention or pre-adjustment for each compression rate, thus maintaining high compression performance while simplifying operation.
Solution Approach 2:
The patent implements preliminary action by pre-training the model to learn optimal quantization parameter settings for various compression scenarios. During inference, the model can quickly select or adjust parameters based on the target bit rate without requiring time-consuming manual optimization, thus maintaining performance while improving ease of operation.
4Device complexity
If a single model is used for variable bit rate compression, then training complexity is reduced, but achieving accurate compression at specific bit rates becomes difficult
Solution Approach 1:
The patent applies dynamics by incorporating dynamic parameter adjustment mechanisms within the single model. The model includes learnable parameters that can be dynamically modified during inference to achieve precise control over the compression bit rate, allowing accurate bit rate targeting without requiring multiple fixed-rate models.
Solution Approach 2:
The patent implements parameter changes by using learnable quantization parameters that can be continuously adjusted to achieve precise bit rate control. The model transforms discrete quantization operations into continuous, differentiable parameter adjustments that can be optimized during training and precisely controlled during inference, achieving accurate bit rate targeting with a single model.
Data Source
AI summary
A variable bit rate generative compression method based on adversarial learning is provided. According to the method, a variance of a feature map of an encoding-decoding fill convolutional network is quantized to train a single generative model to perform variable bit rate compression. The method includes the following implementation steps of: constructing training and testing data sets through an image acquisition device; constructing a generative compression network based on an auto-encoder structure; according to a rate-distortion error calculation unit, alternately training a generative network; according to a target compression rate, calculating a mask threshold; based on a feature map channel redundancy index, calculating a mask; and performing lossless compression and decoding on the mask and the feature map. According to the invention, only a single model is trained, but compression results with different bit rates can be generated, and on a limit compression rate below 0.1 bpp.