Low-power Ship Detection in Satellite Images Using Neuromorphic Hardware (2406.11319v1)

Abstract: Transmitting Earth observation image data from satellites to ground stations incurs significant costs in terms of power and bandwidth. For maritime ship detection, on-board data processing can identify ships and reduce the amount of data sent to the ground. However, most images captured on board contain only bodies of water or land, with the Airbus Ship Detection dataset showing only 22.1\% of images containing ships. We designed a low-power, two-stage system to optimize performance instead of relying on a single complex model. The first stage is a lightweight binary classifier that acts as a gating mechanism to detect the presence of ships. This stage runs on Brainchip's Akida 1.0, which leverages activation sparsity to minimize dynamic power consumption. The second stage employs a YOLOv5 object detection model to identify the location and size of ships. This approach achieves a mean Average Precision (mAP) of 76.9\%, which increases to 79.3\% when evaluated solely on images containing ships, by reducing false positives. Additionally, we calculated that evaluating the full validation set on a NVIDIA Jetson Nano device requires 111.4 kJ of energy. Our two-stage system reduces this energy consumption to 27.3 kJ, which is less than a fourth, demonstrating the efficiency of a heterogeneous computing system.

• The paper proposes a two-stage heterogeneous system combining a low-power neuromorphic classifier (AkidaNet on Akida) with a standard object detector (YOLOv5) for energy-efficient ship detection.
• This approach significantly reduces energy consumption by processing only promising images with the second stage, demonstrating an energy reduction from 111.4 kJ to 27.3 kJ on a validation dataset compared to a single-stage model.
• Implementing such energy-efficient pipelines enables more sustainable and effective satellite Earth observation missions by reducing power demands and data transmission for edge processing.

Low-power Ship Detection in Satellite Images Using Neuromorphic Hardware

This paper presents a novel approach to ship detection in satellite images by leveraging a two-stage system designed to optimize energy efficiency on neuromorphic hardware. The strategic focus is on reducing the power consumption and bandwidth required to process Earth observation data aboard satellites, particularly due to the prevalent challenge where most satellite images do not contain the ships of interest.

Key Contributions

The primary contribution of the paper is a heterogeneous computing system that combines a lightweight binary classifier, AkidaNet 0.5, with a robust object detection model, YOLOv5, running on separate hardware platforms. AkidaNet, executed on Brainchip's neuromorphic processor Akida 1.0, performs the preliminary detection of ships, serving as a gatekeeper to determine whether further in-depth analysis with YOLOv5 is necessary. By utilizing activation sparsity, Akida 1.0 achieves significant power savings, only progressing images with high ship-detection confidence to the next stage.

Technical Achievements

The system architecture showcases several technical achievements:

• Binary Classification Accuracy: The binary classifier attains 97.67% accuracy following quantization-aware training, outperforming prior benchmarks on subsets of the Airbus Ship Detection dataset. This stage prioritizes recall, accepting lower precision, thus optimizing it to reduce false negatives while maintaining energy efficiency.
• Detection Precision: By employing a dedicated ship detection system on images identified by the binary classifier, the system reduces the false-positive rate of YOLOv5, substantiated by an mAP improvement from 76.9% (full set) to 79.3% (ship-containing images only).
• Energy Efficiency: The work demonstrates a noteworthy decrease in energy consumption, achieving a reduction from 111.4 kJ to 27.3 kJ for processing the validation dataset when applied with the NVIDIA Jetson Nano, compared to the single-stage YOLOv5 model.

Significance and Implications

The deployment of such an energy-efficient two-stage pipeline has substantial implications for satellite-based edge computing. The methodology not only addresses power consumption challenges associated with the transmission of vast amounts of mostly irrelevant satellite data but also enhances the operational capacity of space-borne sensors to prioritize and downlink significantly valuable information.

Neuromorphic processors, like Akida, modelled on the human brain's architecture, consolidate precise computing with finely tuned energy efficiency. Their event-based nature, where computation is contingent upon non-zero activations, considerably diminishes both computational load and power use, which is notably advantageous in the context of ship detection amidst predominantly homogeneous backgrounds.

Future Developments

The results indicate potential advancements in edge AI processing via sophisticated neuromorphic chips, including:

• Expanded Application Scope: Future iterations like Akida 2.0, with support for RNNs and transformer models, are poised to further enhance model precision and energy efficiency, promising improvements in other compute-intensive applications on satellites.
• Refinement of Real-Time Processing: Practically, this system could advance real-time spatial awareness in maritime operations, aiding autonomous navigation systems and enhancing responses to environmental incidents.

Conclusion

By integrating neuromorphic processing within a specialized detection workflow, the authors present a pragmatic solution for energy-efficient satellite image analysis. This research underscores the potency of combining low-power neuromorphic computing with high-accuracy object detection models, paving the way for more sustainable and efficacious Earth observation missions. Further optimization through next-gen hardware and nuanced model architectures suggests promising avenues for advancement in space-edge AI applications.