Complex "zero-shot" super-resolution reconstruction algorithm for THz imaging.

Area of Science:

• Terahertz (THz) imaging
• Computational imaging
• Deep learning for image reconstruction

Background:

• Terahertz (THz) imaging systems often suffer from low-resolution limitations.
• Convolutional neural network (CNN) based super-resolution (SR) algorithms typically require extensive standard training datasets.
• A lack of available datasets hinders the application of CNN-based SR in THz imaging.

Purpose of the Study:

• To propose a novel complex "zero-shot" SR (CZSSR) algorithm for THz imaging.
• To address the challenge of reconstructing super-resolved THz images without standard training datasets.
• To enhance the self-adaptability of SR algorithms by leveraging internal image statistics.

Main Methods:

• Development of a five-layer complex CNN model for the CZSSR algorithm.
• Utilizing internal image statistics for image reconstruction, eliminating the need for pre-training.
• Implementation of a zero-shot learning approach tailored for THz SR.

Main Results:

• The CZSSR algorithm demonstrated improved performance compared to traditional zero-shot SR (ZSSR).
• Peak Signal-to-Noise Ratio (SNR) increased by approximately 0.94 dB.
• Mean Squared Error (MSE) decreased by 0.042, and Structural Similarity Index Measure (SSIM) increased by about 40% for measured data.
• The method effectively solved low-resolution problems in THz imaging systems.

Conclusions:

• The proposed CZSSR method successfully overcomes the limitations of standard datasets in CNN-based THz SR.
• The algorithm exhibits strong self-adaptability, making it suitable for diverse THz imaging applications.
• This research holds significant potential for advancing medical imaging and non-destructive detection using THz technology.