Local-enhanced transformer for single-pixel imaging.

Area of Science:

• Optics and photonics
• Computer vision
• Machine learning

Background:

• Deep learning enhances single-pixel imaging (SPI) reconstruction.
• Convolutional neural networks struggle with long-range dependencies in SPI data.
• Transformers excel at long-range dependencies but lack local mechanisms for SPI.

Purpose of the Study:

• To develop a novel deep learning method for high-quality under-sampled SPI.
• To address limitations of existing convolutional and transformer-based SPI reconstruction techniques.
• To improve the modeling of both local and global dependencies in SPI measurements.

Main Methods:

• Proposed a novel local-enhanced transformer architecture for under-sampled SPI.
• Integrated local dependency modeling with the transformer's global dependency capturing ability.
• Utilized optimal binary patterns for efficient and hardware-friendly sampling.

Main Results:

• The local-enhanced transformer effectively models both local and global dependencies in SPI.
• The proposed method demonstrates superior reconstruction quality compared to state-of-the-art SPI techniques.
• Experiments on simulated and real data validate the method's effectiveness.

Conclusions:

• The local-enhanced transformer offers a significant advancement in under-sampled SPI reconstruction.
• The method provides high-quality imaging by effectively combining local and global feature extraction.
• Optimal binary patterns contribute to the efficiency and practicality of the proposed SPI technique.