Abstract
To expedite space exploration on Mars, it is indispensable to develop an efficient Martian image compression method for transmitting images through the constrained Mars-to-Earth communication channel. Although the existing learned compression methods have achieved promising results for natural images from earth, there remain two critical issues that hinder their effectiveness for Martian image compression: 1) They overlook the highly-limited computational resources on Mars; 2) They do not utilize the strong inter-image similarities across Martian images to advance image compression performance. Motivated by our empirical analysis of the strong intra- and inter-image similarities from the perspective of texture, color, and semantics, we propose a reference-based Martian asymmetrical image compression (REMAC) approach, which shifts computational complexity from the encoder to the resource-rich decoder and simultaneously improves compression performance. To leverage inter-image similarities, we propose a reference-guided entropy module and a ref-decoder that utilize useful information from reference images, reducing redundant operations at the encoder and achieving superior compression performance. To exploit intra-image similarities, the ref-decoder adopts a deep, multi-scale architecture with enlarged receptive field size to model long-range spatial dependencies. Additionally, we develop a latent feature recycling mechanism to further alleviate the extreme computational constraints on Mars. Experimental results show that REMAC reduces encoder complexity by 43.51% compared to the state-of-the-art method, while achieving a BD-PSNR gain of 0.2664 dB.
The Challenge of Deep Space Imaging
Transmitting high-fidelity images from Mars to Earth is a massive bottleneck in space exploration. Mars rovers, like NASA’s Perseverance, operate on incredibly limited hardware—often featuring processors and memory (e.g., 256MB DRAM) that are decades behind consumer smartphones.
Existing image compression standards (like HEVC or VVC) and modern AI-based methods are poorly suited for this environment for two reasons. First, there is a Resource Mismatch: they assume the "encoder" (the rover) has significant computing power, or they use "symmetrical" architectures where both encoding and decoding are equally heavy. Second, there is Content Ignorance: Martian landscapes are monotonous and repetitive. Standard algorithms treat every image as a brand-new scene, failing to exploit the fact that a new photo of a crater looks remarkably like previous photos of the same terrain.
Exploiting Martian Uniformity
A key insight of this research is the "Uniformity Advantage." Empirical analysis proves that Martian images have much higher intra-image (within one photo) and inter-image (between different photos) similarity than Earth images. By analyzing the Euclidean distance of Gray-Level Co-occurrence Matrix (GLCM) features, the authors demonstrated that Martian textures are significantly more redundant than those found in diverse Earth-based datasets.
The REMAC Framework
The authors introduce REMAC (Reference-based Martian Asymmetrical Image Compression), which uses three novel strategies to solve these issues:
- Asymmetrical Design: The framework intentionally keeps the R-Encoder (on Mars) lightweight while using a deep, multi-scale R-Decoder (on Earth). The decoder uses "Triple Convolutions" (Tcov) to expand its receptive field, allowing it to reconstruct missing details based on global context.
- Reference-Guided Entropy Module: Instead of compressing an image in a vacuum, REMAC selects previously transmitted "reference images" that look similar to the current one. It uses these references to help predict the probability distribution of the current image's data, which significantly reduces the number of bits needed for transmission.
- Latent Feature Recycling: To save even more power on the rover, the authors developed a "recycling" mechanism. Normally, finding a similar reference image requires a separate, power-hungry search. REMAC reuses the features already extracted during the compression process to perform this search, eliminating redundant calculations.
Key Results and Hardware Realism
REMAC was validated using the MIC (Martian Image Compression) Dataset and raw images from the NASA Mars 2020 Mission (Mastcam-Z). The results demonstrate significant improvements over the state-of-the-art (WACNN):
- Efficiency: REMAC reduces encoder complexity (measured in FLOPs) by 43.51%.
- Quality: It achieves a 0.2664 dB BD-PSNR gain, providing better image quality at the same file size (bitrate).
- Speed: On a CPU (similar to rover hardware), REMAC’s encoding was nearly 3x faster than the VVC standard.
Crucially, the authors confirmed that a block-based version of REMAC consumes only ~207MB of memory, fitting comfortably within the 256MB DRAM limit of actual Mars rovers. This demonstrates that the model is not just theoretically superior but practically deployable for deep space missions.
Future Work
While REMAC is a major step forward, future work will focus on model compression techniques like weight quantization (moving from 32-bit to 8-bit integers) and network pruning. Additionally, the team aims to optimize the code specifically for radiation-hardened processors used in space, which feature different instruction sets than standard Earth-bound CPUs.
Summary updated