Rate-constrained learning-based image compression
作者：
Highlights：
• A non-restricted loss solution can provide rate control in neural image compression.
• Additive uniform noise approximation to quantization can distort the estimated rate.
• Changes to parameters of target-rate loss can overcome training distortion.
• A heuristic that adjusts target-rate loss parameters, avoiding empirical methods.
• A bit allocation analysis that can lead to better rate control optimization methods.
摘要
•A non-restricted loss solution can provide rate control in neural image compression.•Additive uniform noise approximation to quantization can distort the estimated rate.•Changes to parameters of target-rate loss can overcome training distortion.•A heuristic that adjusts target-rate loss parameters, avoiding empirical methods.•A bit allocation analysis that can lead to better rate control optimization methods.
论文关键词：Image coding,Neural networks,Rate–distortion,Rate control,Learning-based compression
论文评审过程：Received 3 May 2021, Revised 11 October 2021, Accepted 31 October 2021, Available online 24 November 2021, Version of Record 11 December 2021.
论文官网地址：https://doi.org/10.1016/j.image.2021.116544
