Deep Lossy Plus Residual Coding for Lossless and Near-Lossless Image Compression

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Lossless and near-lossless image compression are critical for fields like medicine, remote sensing, and scientific research.
• Existing learning-based methods lack unified support for both lossless and near-lossless compression modes.

Purpose of the Study:

• To introduce a novel deep lossy plus residual (DLPR) coding framework.
• To enable unified lossless and near-lossless image compression using a single model.

Main Methods:

• The DLPR framework employs Variational Autoencoders (VAEs) for joint lossy and residual compression.
• Autoregressive context modeling enhances lossless compression of residuals.
• A scalable near-lossless scheme quantizes residuals for variable error bounds (ℓ∞).
• Algorithm parallelization and adaptive entropy coding accelerate the process.

Main Results:

• The DLPR system achieves state-of-the-art performance in both lossless and near-lossless image compression.
• The framework demonstrates competitive coding speeds.
• The proposed scalable scheme efficiently handles variable error bounds.

Conclusions:

• The DLPR coding framework provides a unified and powerful solution for lossless and near-lossless image compression.
• This approach meets the demanding requirements of professional users in technical fields.
• The method offers a significant advancement in image compression technology.