RCodSpace: A Robust Learned Coding Method for Deep Space Visual Transmission

Area of Science:

• Deep space exploration
• Image compression
• Machine learning

Background:

• Deep space missions require transmitting vast image data, but current methods offer limited compression and struggle with transmission errors.
• Learning-based compression excels for natural images but is too complex and error-prone for deep space.

Purpose of the Study:

• To develop a novel image compression and transmission framework for deep space exploration.
• To address the challenges of low-complexity and error resilience in deep space image coding.
• To improve the efficiency and reliability of transmitting Martian image data.

Main Methods:

• Established a new, high-resolution Martian image dataset to guide network design.
• Developed a Martian Vision Adaptive Transformation Module (MVATM) for efficient, low-complexity compression.
• Implemented a two-stage training strategy with Joint Channel Training (JCT) for enhanced error resilience.

Main Results:

• The proposed framework achieves superior rate-distortion-complexity (RDC) trade-offs compared to advanced learning-based models.
• Demonstrated significantly better performance than conventional methods in deep space simulations.
• Hardware deployment validated the method's effectiveness and practicality.

Conclusions:

• The novel framework offers a viable solution for efficient and error-resilient image compression in deep space.
• The developed techniques provide methodological insights for resource-constrained fields.
• This advancement supports future deep space exploration by improving data transmission capabilities.