STanH : Parametric Quantization for Variable Rate Learned Image Compression.

Area of Science:

• Machine Learning
• Image Processing
• Data Compression

Background:

• End-to-end learned image compression models typically require training separate encoder-decoder pairs for each target bitrate (λ).
• This approach leads to significant storage and computational overhead, necessitating multiple models for different compression levels.
• Optimizing the rate-distortion trade-off (R + λD) is crucial for efficient image compression.

Purpose of the Study:

• To develop a novel method for achieving variable rate image compression using a single, adaptable model.
• To reduce the storage and training complexity associated with traditional learned image compression techniques.
• To enable flexible control over the rate-distortion trade-off without retraining.

Main Methods:

• Proposes STanH, a differentiable quantizer based on a parametric sum of hyperbolic tangents, relaxing the step-wise quantization.
• Implements STanH as a learnable activation layer within a pre-trained fixed-rate image compression model.
• Refines the model with the STanH layer to achieve various target bitrates.

Main Results:

• The proposed STanH method enables variable rate coding with performance comparable to state-of-the-art methods.
• Demonstrates significant savings in deployment ease, training time, and storage costs.
• The differentiable nature of STanH allows for seamless integration and adaptation.

Conclusions:

• STanH offers an efficient and flexible solution for learned image compression, addressing the limitations of fixed-rate models.
• This approach significantly lowers the barrier to deploying learned image compression technologies.
• The method provides a practical pathway towards high-efficiency, adaptable image compression systems.