Generalized Gaussian Model for Learned Image Compression.

Area of Science:

• Computer Vision
• Machine Learning
• Information Theory

Background:

• Probabilistic models are crucial for latent variable distribution in learned image compression.
• The standard Gaussian model offers simplicity but limited flexibility.
• Gaussian mixture models provide better fit but increase complexity.

Purpose of the Study:

• To develop a more flexible probabilistic model for latent variables in learned image compression.
• To balance compression performance and model complexity.
• To improve the accuracy and efficiency of learned image compression techniques.

Main Methods:

• Extended the standard Gaussian model to a generalized Gaussian model with an additional shape parameter.
• Introduced improved training strategies: -dependent lower bounds for scale parameters and gradient rectification.
• Evaluated the proposed model on various learned image compression networks.

Main Results:

• The generalized Gaussian model demonstrated superior performance compared to Gaussian and Gaussian mixture models.
• The enhanced training methods effectively reduced the train-test mismatch.
• Achieved better compression performance across diverse image compression networks.

Conclusions:

• The generalized Gaussian model offers a flexible and effective approach for latent distribution modeling in learned image compression.
• The proposed training enhancements significantly boost model performance.
• This work advances the state-of-the-art in learned image compression by improving probabilistic modeling.