Multi-Band Spectral Fusion Terahertz Deep Learning Computed Tomography

We present a novel method to visualize 3D object details through a multi-band spectral fusion deep learning approach. This method, multi-dominant-band spatio-spectral fusion network (MS-Unet), incorporates rich information from multi-frequency band spectral data. The reconstructed terahertz tomographic images significantly outperform conventional methods, enhancing 8.97 dB in PSNR, boosting SSIM from 0.05 to 0.70, and improving LPIPS from 0.36 to 0.18.