|
Structured illumination(SI) has a wide range of applications in computational imaging in the form of encoding and recomposing image signals. In this work, we explore the applications and advantages of deep-learning techniques in SI imaging problems. (1) In single-pixel imaging(SPI), where SI encodes the target into a sequence of bucket signals, the recovery suffers from loss of imaging quality due to various noises. We propose an unsupervised deep-learning (UnDL) based anti-noise approach, which outperforms conventional single-pixel imaging methods considerably in reconstructing targets against noise. (2) In blind ghost imaging, we propose two hybrid quantum-classical machine learning algorithms and a physical-inspired patch strategy, leveraging quantum machine learning to restore high-quality images where classical machine learning fails. (3) In structured illumination microscopy(SIM), the illumination encodes high-frequency details into the passband of the objective. Existing optimization-based decoding algorithms are sensitive to noise, while learning-based methods have faithfulness issues. We propose a physics-informed deep learning approach, where the re-parameterized network outperforms its counterparts in terms of noise robustness and recovery faithfulness. In general, the proposed deep-learning-based algorithms solve the inverse SI decoding problem by leveraging inherent priors of neural networks. The proposed framework and idea have the potential to be applied in other scenarios with SI and other computational imaging problems.
|
