Terahertz image super-resolution based on a complex convolutional neural network.

Area of Science:

• Physics
• Imaging Science
• Artificial Intelligence

Background:

• Terahertz (THz) imaging offers unique capabilities for medical and industrial applications.
• A significant challenge in THz imaging is its inherent low resolution due to the long wavelength of THz radiation.
• Conventional convolutional neural networks (CNNs) improve optical image resolution but primarily process real-valued intensity data.

Purpose of the Study:

• To extend the application of CNNs to the complex domain for THz imaging.
• To leverage both amplitude and phase information available in THz data.
• To overcome the resolution limitations in THz imaging and enhance image quality.

Main Methods:

• Developed a complex convolution neural network (CCNN) that operates in the complex number domain.
• Utilized the wave nature of THz light to process both amplitude and phase information.
• Trained the CCNN using the MNIST dataset for image reconstruction tasks.

Main Results:

• Achieved a resolution of 0.4 times the beam size, with potential for half-wavelength resolution.
• Demonstrated a 27.8% increase in image contrast compared to standard CNNs.
• Successfully recovered phase information, which is not possible with standard CNNs.
• Showcased superior generalization capability of the CCNN over the CNN.

Conclusions:

• The CCNN is the first successful complex-valued neural network for THz imaging, significantly improving resolution.
• This approach mitigates the need for critical optical components and extensive system fine-tuning.
• The CCNN offers a powerful solution to the low-resolution bottleneck in THz imaging, benefiting applications like biomedical imaging and non-destructive testing.