DeepFSI Improves Feature Detection in Photon-Counting with Poisson Noise

Photon-counting sensors, crucial for low-light imaging, increasingly suffer from a type of noise called Poisson noise, which limits image clarity. Conventional Feature-Specific Imaging (FSI) techniques, however, are designed for a different type of noise, leaving them underperforming in these conditions. Yizhou Lu from the University of Wisconsin-Madison and colleagues address this challenge with DeepFSI, a new framework that uses artificial intelligence to optimise how images are captured. DeepFSI learns the best way to measure image features directly under realistic noise conditions, resulting in significantly improved image quality and robustness, particularly when light is scarce. This advance promises to enhance imaging in photon-limited applications, such as medical diagnostics and astronomy, where capturing clear images with minimal light is essential.

DeepFSI effectively addresses limitations within traditional feature-specific imaging (FSI) by allowing a deep neural network to learn globally optimal measurement masks. This is achieved through direct gradient computation under realistic conditions involving both Poisson and additive noise. Simulations demonstrate that DeepFSI achieves superior feature fidelity and task performance when compared to conventional FSI methods employing predefined masks, particularly in environments dominated by Poisson noise. Furthermore, DeepFSI exhibits enhanced robustness to variations in design choices and maintains strong performance even under additive Gaussian noise, representing a significant advancement for noise-robust computational imaging in photon-limited applications.

Deep Learning Optimizes Imaging System Design

Researchers have developed a new imaging framework, DeepFSI, that significantly improves performance in low-light conditions where photon noise is a major limitation. DeepFSI overcomes this limitation by allowing a deep learning network to learn optimal measurement patterns directly from the data, even under realistic, noisy conditions. This approach represents a fundamental shift in how coding-based imaging systems are designed and optimized.

The core innovation lies in the system’s ability to learn measurement patterns tailored to the specific features being imaged, rather than relying on pre-defined, static patterns. By incorporating the effects of photon noise directly into the training process, DeepFSI learns to extract information more effectively, leading to substantial gains in image fidelity and task performance. Unlike previous methods that required precise knowledge of noise levels and the number of features, DeepFSI demonstrates remarkable robustness, maintaining high performance even when these parameters are inaccurate or change. This adaptability broadens the applicability of the technology to a wider range of real-world scenarios.

Deep Learning Optimizes Low-Light Imaging Performance

Testing on a single-pixel camera, a system well-suited for evaluating coding strategies, reveals that DeepFSI consistently outperforms conventional FSI in photon-noise-dominant environments. The improvement isn’t merely incremental; the system learns to compensate for the inherent randomness of photon detection, extracting meaningful signals where traditional methods fail. Furthermore, the research demonstrates that DeepFSI’s performance is not limited to low-light scenarios; it also performs well under more conventional noise conditions, highlighting its versatility and potential for broader adoption in various imaging applications. This advancement promises to enhance the practicality of FSI techniques and unlock new possibilities for imaging in challenging conditions.

Deep Learning Optimises Photon Counting Imaging

DeepFSI introduces a new approach to feature-specific imaging, addressing limitations of conventional methods when dealing with the prevalent Poisson noise in modern photon-counting sensors. The research team developed a deep learning framework that optimises measurement masks, allowing the system to learn globally optimal settings directly under realistic noise conditions. Simulations demonstrate that DeepFSI achieves superior feature fidelity and task performance compared to traditional feature-specific imaging, particularly in environments dominated by Poisson noise, and exhibits robustness to variations in design choices. This advancement extends the applicability of photon-counting sensors, offering improved performance in low-light conditions and enhancing practicality for real-world implementation.

The team acknowledges a limitation in their work, noting that the model relies on an approximation of Poisson noise which may not perform optimally at extremely low light levels. Future research could focus on developing a more accurate reparameterisation model specifically tailored for these challenging scenarios. The findings suggest that hardware optimisation, combined with noise-aware training, is crucial for achieving peak performance in future camera and vision systems.