Deep learning based projector defocus compensation in single-pixel imaging.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Artificial Intelligence

Background:

• Fourier single-pixel imaging (FSI) reconstructs images using projected patterns and a single-pixel detector.
• Projector defocusing blurs patterns, degrading image quality in FSI and other single-pixel imaging (SPI) schemes.
• Existing methods struggle to effectively compensate for projector defocusing.

Purpose of the Study:

• To introduce a novel deep learning (DL) approach for compensating projector defocusing in FSI.
• To develop a fast, adaptive, and scalable solution for improving image reconstruction quality under defocus conditions.
• To demonstrate the method's robustness and superiority over conventional techniques.

Main Methods:

• A deep convolutional neural network (DCNN) was trained on a large dataset of FSI images reconstructed with varying defocus parameters.
• The DCNN was further trained using experimental data to enhance robustness against system-specific biases.
• The proposed DL model learns to counteract the blurring effects caused by projector defocus.

Main Results:

• The DL method successfully reconstructed high-quality FSI images even at high levels of projector defocus.
• Experimental results validated the method's efficacy and robustness.
• Comparative analysis showed the proposed approach outperforms conventional FSI and existing defocus rectification methods.

Conclusions:

• The proposed deep learning method effectively compensates for projector defocusing in FSI, enabling high-quality image reconstruction.
• This approach offers a scalable and adaptive solution applicable to various SPI techniques affected by optical anomalies.
• The work highlights the potential of DL for correcting optical aberrations in imaging systems.