Neuromorphic Camera Denoising Using Graph Neural Network-Driven Transformers.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Neuromorphic vision offers advantages like low power and high speed.
• Neuromorphic cameras are susceptible to significant measurement noise.
• This noise degrades the performance of perception and navigation algorithms.

Purpose of the Study:

• To propose a novel noise filtration algorithm for neuromorphic cameras.
• To eliminate noise events that do not represent actual scene changes.
• To improve the reliability of neuromorphic event-based systems.

Main Methods:

• A graph neural network (GNN)-driven transformer algorithm (GNN-Transformer) classifies events as real or noise.
• An EventConv message-passing framework captures spatiotemporal correlations while maintaining event asynchronicity.
• A Known-Object Ground-Truth Labeling (KoGTL) approach generates labeled datasets under varied lighting.

Main Results:

• The GNN-Transformer algorithm achieved at least 8.8% higher filtration accuracy than state-of-the-art methods on unseen datasets.
• The algorithm demonstrated superior noise elimination while preserving meaningful events.
• Tests on ETH Zürich Color-DAVIS346 datasets confirmed generalization capabilities across illumination and motion variations.

Conclusions:

• The proposed GNN-Transformer algorithm effectively filters noise in neuromorphic vision.
• This noise reduction significantly enhances the performance of event-based perception and navigation.
• The method shows strong generalization, making it suitable for diverse real-world applications.