Work

Data-efficient self-supervised 3DCNN for real-time terahertz image denoising and enhancement

Year: 2026

Type: article

Abstract: Terahertz (THz) imaging is recognized for its distinctive capability in noninvasive detection across various applications, including biomedical diagnostics, security inspection, and material characterization. However, THz images often suffer from significant noise interference due to inherent limitations such as low detector sensitivity, diffraction effects, and inter-pixel crosstalk, severely compromising the interpretability and accuracy of resulting images. Traditional supervised deep learning methods require large annotated datasets, challenging in THz imaging due to limited availability and high acquisition costs. To address these limitations, this study proposes an efficient self-supervised temporal denoising framework specifically designed for real-time THz imaging. The framework comprises a novel lightweight three-dimensional convolutional neural network (3D-TDCNN) and a highfrequency guided self-supervised temporal denoising model (HF2HF), leveraging three-dimensional discrete wavelet transforms (3D-DWT) to specifically target noise-rich high-frequency sub-bands. Extensive experiments using simulated and real-world THz image datasets demonstrate that the proposed method achieves superior performance compared to state-of-the-art supervised and self-supervised techniques. Our approach significantly improves image clarity and preserves critical structural details while requiring minimal training data, offering a robust and practical solution for high-fidelity, real-time THz imaging.Preprint submitted to Expert Systems with Applications June 26, 2025

Source: Optics & Laser Technology

DOI: 10.1016/j.optlastec.2026.114656