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

• A self-supervised 3D CNN is designed for terahertz video denoising with limited data. • High-frequency guidance using 3D wavelet transform preserves temporal and spatial details. • The network enables real-time denoising of noisy terahertz video streams. • Experimental results show superior performance over existing supervised and blind methods. • The framework supports robust THz imaging in low-SNR and label-scarce environments. Terahertz (THz) focal-plane array (FPA) imaging is attractive for real-time, non-invasive inspection, but practical systems suffer from low signal-to-noise ratio, scarce clean labels and stringent throughput constraints. In this work, we propose a data-efficient self-supervised temporal denoising framework tailored to real-time THz video streams. The method combines a lightweight three-dimensional THz denoising convolutional neural network (CNN) with a high-frequency–guided self-supervised module operating in the 3D wavelet domain. A 3D discrete wavelet transform decomposes each x–y–t volume into sub-bands, and the network selectively denoises only the temporal–spatial high-frequency components while preserving low-frequency structure. Supervision is obtained from odd–even frame pairs generated by sliding windows over noisy sequences, eliminating the need for ground-truth labels. The framework is integrated with a custom 340 GHz THz array camera using a 64 × 64 FPA and achieves real-time inference at 56 fps on a commodity GPU, exceeding the sensor readout rate and enabling in-loop denoising. Experiments on physically grounded synthetic videos and 17 real terahertz sequences show consistent gains over classical and learning-based baselines, yielding higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) in supervised comparisons, and better No-Reference Quality Metric (NRQM), Neural Image Assessment (NIMA), Naturalness Image Quality Evaluator (NIQE), and Perceptual Index (PI) on real data. The proposed framework therefore provides a practical, label-efficient and implementation-friendly solution for high-fidelity, real-time THz video denoising and enhancement in non-destructive testing and related applications.