Improving NIR single-pixel imaging: using deep image prior and GANs.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Single-pixel imaging (SPI) is valuable in low light or with limited spectral cameras, especially in the near-infrared (NIR) range (850-1550 nm).
• Traditional SPI requires extensive datasets for image enhancement, limiting its application in specific spectral bands.
• Deep learning methods offer potential for image super-resolution but often need large, paired datasets.

Purpose of the Study:

• To introduce a hybrid deep image prior (DIP) and generative adversarial network (GAN) approach for single-pixel imaging (SPI) super-resolution.
• To reduce the reliance on large, direct SPI image datasets for image quality enhancement.
• To improve the resolution of SPI, particularly in the challenging near-infrared (NIR) spectral range.

Main Methods:

• A hybrid model combining Deep Image Prior (DIP) and Generative Adversarial Networks (GANs) was developed.
• Unsupervised image super-resolution techniques based on DIP were employed to minimize dataset requirements.
• Neural network architectures, including UNet and GANs, were enhanced and tested across four configurations.

Main Results:

• The proposed hybrid DIP-GAN method successfully improved the resolution of single-pixel images.
• Numerical and experimental evidence validated the effectiveness of the approach.
• The method demonstrated enhanced image quality in specific NIR bands without extensive training data.

Conclusions:

• The hybrid DIP-GAN approach offers an effective solution for enhancing SPI resolution, especially in the NIR spectrum.
• This method simplifies the process of improving SPI image quality by reducing data dependency.
• The findings support the broader applicability of SPI in challenging imaging conditions.