Single-pixel imaging using a recurrent neural network combined with convolutional layers.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Single-pixel imaging offers advantages like high-speed acquisition and system miniaturization, but suffers from low image quality due to undersampling.
• Conventional deep learning methods for image reconstruction require substantial memory for internal parameters, limiting practical applications.

Purpose of the Study:

• To develop an efficient deep learning-based single-pixel imaging method with reduced computational complexity.
• To improve image reconstruction quality and robustness against noise in undersampled single-pixel imaging systems.

Main Methods:

• Implementation of a recurrent neural network (RNN) architecture tailored for single-pixel imaging data.
• Training and validation of the RNN model using simulated and experimental undersampled single-pixel imaging data.

Main Results:

• The proposed RNN-based method significantly reduced the number of internal parameters compared to conventional deep learning models.
• Reconstructed images exhibited higher quality and improved detail preservation, even under severe undersampling conditions.
• The model demonstrated notable robustness to various noise types, maintaining performance in challenging imaging scenarios.

Conclusions:

• Recurrent neural networks provide an effective solution for enhancing single-pixel imaging quality while minimizing memory footprint.
• This approach advances the feasibility of single-pixel imaging for applications requiring high performance and efficiency.
• The proposed method represents a significant step towards overcoming the limitations of undersampling in single-pixel imaging systems.