Adaptive terahertz image super-resolution with adjustable convolutional neural network.

Area of Science:

• Optics and Photonics
• Image Processing
• Terahertz Technology

Background:

• Real-aperture-scanning imaging in terahertz (THz) frequencies suffers from inherently low spatial resolution.
• Degradation in THz images is dependent on the imaging range and system's focused beam distribution.

Purpose of the Study:

• To propose an accommodative super-resolution framework for enhancing spatial resolution in THz images.
• To address the range-dependent super-resolution challenge in THz imaging systems.

Main Methods:

• Developed a 3D degradation model incorporating focused THz beam distribution to link imaging range and restoration level.
• Introduced an adjustable convolutional neural network (CNN) capable of producing arbitrary super-resolution levels by tuning an interpolation parameter.
• Implemented a system where the interpolation coefficient is selected based on the measured imaging range for optimal image restoration.

Main Results:

• The proposed framework effectively enhances the spatial resolution of THz images.
• Demonstrated superior super-resolution effects on both simulated and real terahertz data acquired by a 160–220 GHz imager.
• The adjustable CNN achieved desired super-resolution levels without requiring additional training for intermediate ranges.

Conclusions:

• The accommodative super-resolution framework offers a flexible and effective solution for improving THz image quality.
• The range-dependent super-resolution approach using an adjustable CNN provides significant advantages for real-aperture THz imaging.
• This method advances the capabilities of THz imaging systems by enabling high-resolution imaging across various operational scenarios.