使用Bits-Back Coding於無失真變分影像壓縮之演算法實現與性能分析

Abstract

本研究著重於無失真的變分影像壓縮，採用了變分自動編碼器（VAE）模型進行無失真壓縮，同時搭配Bits-Back Coding演算法，該演算法天生適用於VAE模型。在本篇研究中，我們提供了對Bits-Back Coding的詳細介紹，並整合了幾篇結合Bits-Back Coding和VAE的相關論文的程式碼，進行了實驗比較。最終，針對無失真的變分影像壓縮，我們提供了在不同限制條件和情境下的建議，以協助選擇合適的模型或工具套件。這篇論文的貢獻在於深入探討了神經網路無失真資料壓縮的其中一種方法和技術，並提供了實用的指導方針，期望能對進一步的相關研究和拓展神經網路資料壓縮技術的實際應用效果有所助益。

This study focuses on lossless variational image compression using the VAE model, combined with the Bits-Back Coding algorithm, which is naturally suited for VAE models. We comprehensively introduce Bits-Back Coding and integrate code from several related papers that combine Bits-Back Coding with VAE for experimental comparison. Ultimately, we offer recommendations for lossless variational image compression under different constraints and scenarios to assist in choosing suitable models or toolkits. The contribution of this paper lies in its in-depth exploration of one method and technique for neural network-based lossless data compression, along with practical guidelines that can facilitate further research and the practical application of neural data compression technology.