Terahertz image enhancement based on a multiscale feature extraction network.

Area of Science:

• Optics and Photonics
• Image Processing
• Artificial Intelligence

Background:

• Terahertz (THz) waves offer significant potential across military, industrial, and biomedical applications.
• Terahertz time-domain spectroscopy (THz-TDS) imaging faces challenges like diffraction limits, atmospheric absorption, and noise, impacting image quality.
• Existing THz imaging techniques struggle with resolution, contrast, and noise reduction.

Purpose of the Study:

• To enhance the quality of terahertz images, specifically improving resolution and denoising.
• To develop a novel deep learning-based super-resolution network for THz imaging.
• To evaluate the proposed algorithm's performance against established methods for THz image enhancement.

Main Methods:

• A generative adversarial network (GAN) structure was employed for deep learning-based THz image enhancement.
• The network integrates encoder-decoder concepts and a pyramid pooling residual dense block for feature extraction.
• The super-resolution network was applied to terahertz images of deformed metal samples.

Main Results:

• The proposed algorithm demonstrated superior performance in improving image resolution compared to bicubic, SRGAN, and RDN methods.
• Effective noise removal was achieved while preserving essential high-frequency details in the terahertz images.
• The method successfully avoided the introduction of unwanted high-frequency artifacts.

Conclusions:

• The developed deep learning algorithm significantly enhances terahertz image quality, offering improved resolution and denoising.
• This approach effectively addresses key challenges in terahertz imaging, paving the way for advanced applications.
• The method preserves fine details, making it valuable for analyzing materials like deformed metals using THz-TDS.