Self-supervised deep-learning for efficient denoising of terahertz images measured with THz-TDS system

Abstract

This paper introduces self-supervised deep-learning approach for efficiently denoising terahertz (THz) images acquired using a THz time-domain spectroscopy (THz-TDS) imaging system, addressing the critical challenge of noise reduction in THz imaging, which often limits its practical applications. To overcome the limitations of conventional denoising methods—such as the need for clean image pairs or precise noise modeling—we adopted the Noise2Noise methodology, which enables effective training using only noisy images. To optimize the training process, we applied zero-padding to the THz-TDS time-domain data and determined the optimal number of training images and zero-padding ratio. Quantitative evaluation using THz images of the Positive 1951 USAF Test Target demonstrates significant improvement: the noise levels of the 1.20 THz images, initially ranging from 16.87 to 4.55 for signal-to-noise ratios (SNRs) of 12.5 dB to 33.1 dB, were reduced significantly to 0.61 to 0.36, corresponding to denoising rates of 96% to 92%. These results demonstrate the model’s effectiveness in handling THz images with varying noise levels. Furthermore, extensive experimental results validate the proposed model’s consistent denoising efficiency, achieving superior performance across diverse noise environments and image structures. This provides a strong foundation for its application in real-world THz imaging systems. Additionally, we analyzed the versatility of the model by training it on USAF THz images with linear structures and various noise levels and frequencies, and then testing it on Samjoko (triple-legged crow) bookmark THz images with curved structures. The model consistently exhibited excellent denoising performance, demonstrating its robustness and adaptability to different image shapes, noise levels, and frequencies. The success of our deep-learning model in maintaining consistent denoising capability across diverse conditions establishes its potential for real-world THz imaging applications. These results prove that laboratory-trained models can achieve superior performance in various practical settings.