Scalable Face Image Coding via StyleGAN Prior: Toward Compression for Human-Machine Collaborative Vision.

Area of Science:

• Computer Vision
• Image Compression
• Machine Learning

Background:

• The increasing volume of visual data and advancements in machine vision necessitate efficient methods for data representation and delivery.
• Current compression techniques struggle to balance the needs of both human perception and machine analysis, especially at scale.

Purpose of the Study:

• To develop an efficient scalable coding paradigm for human-machine collaborative vision.
• To leverage advanced generative priors for creating hierarchical representations of visual data.
• To optimize compression for both machine analysis performance and human perceptual quality.

Main Methods:

• Exploiting the StyleGAN generative prior to learn three-layered hierarchical representations (basic, middle, enhanced).
• Proposing a layer-wise scalable entropy transformer to minimize inter-layer redundancy.
• Jointly optimizing the scheme using a multi-task scalable rate-distortion objective.

Main Results:

• The proposed paradigm demonstrates superior performance in face image compression compared to the Versatile Video Coding (VVC) standard.
• Achieved significant improvements in both machine analysis and human perception at extremely low bitrates (< 0.01 bpp).
• Validated the feasibility of hierarchical representations for human-machine collaborative compression.

Conclusions:

• Hierarchical representations derived from generative priors offer an effective approach for scalable visual data coding.
• The proposed method provides a new paradigm for human-machine collaborative compression, balancing efficiency, machine analysis, and human perception.
• This work offers valuable insights for future research in efficient visual data representation and compression.