Self-Supervised Deep Learning for THz-TDS Image Denoising

Abstract

This study presents a self-supervised deep learning methodology for efficient denoising of terahertz (THz) images acquired through THz time-domain spectroscopy (THz-TDS) systems. Unlike traditional approaches that require paired clean and noisy images, the proposed method leverages the Noise2Noise framework to enable effective model training using only noisy data. Quantitative evaluations on 1.20 THz images demonstrate a significant noise level reduction from 16.87 to 0.61, corresponding to a 96% improvement across diverse signal-to-noise ratio (SNR) conditions. Furthermore, the model exhibits strong generalization capabilities across various frequency bands and structural characteristics, consistently achieving noise reductions of 74% to 96% throughout the 1.00-2.00 THz spectrum. © 2025 IEEE.