A lightweight model for perceptual image compression via implicit priors.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Perceptual image compression offers superior visual quality at low bitrates compared to traditional methods.
• Existing approaches often rely on explicit semantic features, increasing model complexity and limiting mobile application.

Purpose of the Study:

• To develop a lightweight perceptual image compression method that reduces model complexity while maintaining high visual quality.
• To leverage implicit semantic priors for efficient image compression on resource-constrained devices.

Main Methods:

• Proposed a lightweight perceptual image compression method using Implicit Semantic Priors (ICISP).
• Developed an enhanced visual state space block for exploiting spatial dependencies and reducing redundancy.
• Introduced a frequency decomposition modulation block for adaptive frequency information preservation.
• Utilized a semantic-informed discriminator with implicit priors from a DINOv2 encoder for enhanced perceptual quality.

Main Results:

• The ICISP method demonstrates competitive compression performance on standard benchmarks.
• Achieved significantly fewer network parameters and floating-point operations compared to state-of-the-art methods.
• The approach effectively balances compression efficiency and perceptual quality.

Conclusions:

• ICISP offers a practical solution for high-quality perceptual image compression on mobile devices.
• Implicit semantic priors provide an efficient alternative to explicit features for reducing model complexity.
• The proposed method advances the field of efficient and visually optimized image compression.