Reproduce “Unified Multivariate Gaussian Mixture for Efficient Neural Image Compression” on OpenResearch

About this paper

Xiaosu Zhu, Jingkuan Song, Lianli Gao, Feng Zheng, Heng Tao Shen

This paper presents a neural image compression framework that utilizes a unified multivariate Gaussian mixture as a vectorized prior for latent variables. The method achieves state-of-the-art rate-distortion performance while enabling significantly faster encoding and decoding speeds by eliminating the need for computationally expensive context models.

Key insights

• Treating latent variables as N-dimensional vectors and modeling their distribution with a multivariate Gaussian mixture can effectively capture richer statistical dependencies than univariate priors.
• A hierarchical, cascaded estimation strategy, combined with a multi-codebook structure, can replace traditional context models to achieve high R-D performance with significantly improved computational efficiency.
• Probabilistic Vector Quantization, when paired with Euclidean distance and a multi-codebook design, facilitates robust training and learning of expressive codebooks, preventing local optima and unused codewords.

Summary by alphaXiv.