Machines Serve Human: A Novel Variable Human-Machine Collaborative Compression Framework

Area of Science:

• Computer Vision
• Image and Video Compression
• Artificial Intelligence

Background:

• Existing human-machine collaborative compression methods rely on human-vision pipelines, which are inefficient for machine vision tasks.
• Machine vision requires less data than human vision, focusing on core image regions.
• Current methods struggle with complexity and bit-rates when integrating machine vision requirements.

Purpose of the Study:

• To develop a novel collaborative compression method prioritizing machine vision.
• To improve efficiency and performance for both human and machine vision data reduction.
• To establish a new paradigm for human-machine collaborative compression.

Main Methods:

• Introduced a novel collaborative compression method based on machine-vision-oriented compression, named Diff-FCHM.
• Developed a plug-and-play variable bit-rate strategy tailored for machine vision.
• Progressively aggregated semantics from machine-vision compression and utilized diffusion priors for human vision detail restoration.

Main Results:

• Diff-FCHM demonstrated consistently superior performance in both machine-vision and human-vision compression.
• Achieved remarkable margins in performance compared to existing methods.
• Validated the effectiveness of the machine-vision-first approach for collaborative compression.

Conclusions:

• The proposed Diff-FCHM method offers a significant advancement in human-machine collaborative compression.
• Machine vision as the basis for collaborative compression leads to enhanced efficiency and performance.
• The novel approach successfully balances the needs of both human and machine vision for data reduction.