Rethinking Satellite Image Restoration for Onboard AI: A Lightweight Learning-Based Approach
Satellite image restoration is a crucial process aimed at enhancing image quality by mitigating degradations such as noise and blur that arise during the imaging process. This restoration serves as a fundamental preprocessing step, significantly influencing both ground-based product generation and the development of emerging onboard artificial intelligence (AI) applications.
Traditional restoration methods typically rely on sequential physical models, which are often computationally intensive and slow. This makes them less suitable for onboard environments where processing power and time are at a premium. In response to these challenges, a research team has introduced a novel model known as ConvBEERS: Convolutional Board-ready Embedded and Efficient Restoration for Space.
Introducing ConvBEERS
The ConvBEERS model is designed to explore the potential of a lightweight and non-generative residual convolutional network, trained on simulated satellite data. The primary goal is to determine whether this innovative approach can match or even exceed the performance of traditional ground-processing restoration pipelines under a variety of operating conditions.
Key Findings
- Image Quality Improvement: Experiments conducted using both simulated datasets and actual Pleiades-HR imagery revealed that the ConvBEERS model achieves a notable +6.9dB Peak Signal-to-Noise Ratio (PSNR) improvement in image quality.
- Enhanced Object Detection: Evaluation of the model’s performance in a downstream object detection task indicated a significant enhancement, resulting in up to +5.1% mean Average Precision (mAP@50) improvement.
- Practical Feasibility: The successful deployment of ConvBEERS on a Xilinx Versal VCK190 FPGA underscores its practical viability for satellite onboard processing, achieving an impressive ~41x reduction in latency compared to the traditional restoration pipeline.
Conclusion
The results of this study highlight the effectiveness of utilizing lightweight convolutional neural networks (CNNs) to attain competitive restoration quality while addressing the real-world constraints faced by spaceborne systems. The findings not only pave the way for advancements in satellite image processing but also open new avenues for integrating AI applications directly onboard satellites.
As the demand for high-quality satellite imagery continues to grow, the ConvBEERS model represents a significant step forward in optimizing image restoration processes, thereby enhancing the capabilities of future satellite missions.