Related Experiment Video
Updated: Aug 15, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
THz-Net: a multi-scale self-attention residual network for cross-system terahertz image reconstruction
Optics Express
|August 14, 2026
Summary
We developed THz-Net, a novel deep learning model for reconstructing high-resolution terahertz (THz) images. This advanced network significantly improves image quality across different systems, overcoming limitations of existing methods.
Area of Science:
- Terahertz (THz) imaging
- Image reconstruction
- Deep learning for imaging
Background:
- Terahertz (THz) imaging is crucial for biomedical diagnosis and non-destructive testing.
- THz images suffer from blur, low contrast, and loss of detail due to diffraction.
- Current reconstruction methods struggle with feature extraction, multi-scale integration, and cross-system generalization.
Purpose of the Study:
- To develop a robust THz image reconstruction method overcoming existing limitations.
- To enhance high-frequency feature extraction and multi-scale information integration.
- To achieve consistent performance across diverse THz imaging systems.
Main Methods:
- Proposed THz-Net, a multi-scale self-attention residual network.
- Utilized an adaptive residual backbone for contrast preservation.
- Implemented multi-scale feature extraction and a dual-attention module (channel and spatial).
Main Results:
- THz-Net achieved an average PSNR of 37.41dB and SSIM of 0.97.
- Outperformed the best baseline by 3.93dB (PSNR) and 0.08 (SSIM).
- Demonstrated consistent reconstruction quality across three different imaging systems.
Conclusions:
- THz-Net effectively reconstructs high-resolution THz images with enhanced detail and contrast.
- The proposed network offers superior performance and generalization capabilities for cross-system THz imaging.
- This advancement holds significant potential for applications in biomedical and non-destructive testing fields.