Deep feature-specific imaging.

Area of Science:

• Computational imaging
• Optical-electronic systems
• Sensor technology

Background:

• Photon-counting sensors are susceptible to Poisson noise.
• Conventional feature-specific imaging (FSI) using principal component analysis (PCA) is optimized for Gaussian noise, not Poisson noise.
• This leads to suboptimal performance in modern imaging applications.

Purpose of the Study:

• To introduce DeepFSI, a novel end-to-end optical-electronic framework.
• To enable deep neural networks to learn optimal measurement masks for noisy conditions.
• To improve performance and robustness in photon-limited imaging.

Main Methods:

• Developed a novel end-to-end optical-electronic framework named DeepFSI.
• Unfrozen PCA-derived masks to allow a deep neural network to learn globally optimal measurement masks.
• Computed gradients directly under realistic Poisson and additive noise conditions.

Main Results:

• DeepFSI demonstrated improved classification accuracy compared to PCA-based FSI.
• Achieved stronger transfer robustness across varying photon budgets, especially in Poisson-dominant environments.
• Showed enhanced robustness to design choices and performed well under additive Gaussian noise.

Conclusions:

• DeepFSI represents a significant advance for noise-robust computational imaging.
• The framework is particularly beneficial for photon-limited applications dominated by Poisson noise.
• DeepFSI offers superior performance and robustness over conventional PCA-based FSI.