Shao Jingzhu, Tang Ping, Zhao Xiangyu, Xu Borui, Chen Bo, Wang Kai, Xu Gangyi, Cao Juncheng, Wu Chongzhao
Center for Biophotonics, Institute of Medical Robotics, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China.
Key Laboratory of Infrared Imaging Materials and Detectors, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai, China.
iScience. 2025 Aug 7;28(9):113278. doi: 10.1016/j.isci.2025.113278. eCollection 2025 Sep 19.
Ptychography is a phase imaging technique that leverages intensity images obtained by translating objects across an illumination beam. Deep learning has demonstrated promising potential in solving inverse problems, offering effective solutions for phase retrieval. However, obtaining substantial amounts of labeled data in the terahertz (THz) bands for pretraining the neural networks is very challenging, thereby limiting the generalization ability of the networks. Here, we propose an untrained physics-driven neural network (UPNN) to facilitate the high-quality reconstruction of THz ptychography. UPNN does not require pretraining and takes only a single dataset as input, which guides the inference process of neural networks by integrating real-world physical models. Therefore, UPNN is more flexible to different types of imaging objects and hardware than the pretraining approach. Moreover, the experimental results and simulated analysis prove that UPNN outperforms conventional iterative ptychography algorithms in terms of image quality, lateral resolution, and robustness.
叠层成像术是一种相位成像技术,它利用通过在照明光束上平移物体而获得的强度图像。深度学习在解决逆问题方面已展现出颇具前景的潜力,为相位恢复提供了有效的解决方案。然而,在太赫兹(THz)频段获取大量用于神经网络预训练的标记数据极具挑战性,从而限制了网络的泛化能力。在此,我们提出一种未经训练的物理驱动神经网络(UPNN),以促进太赫兹叠层成像术的高质量重建。UPNN无需预训练,仅以单个数据集作为输入,通过整合现实世界的物理模型来引导神经网络的推理过程。因此,与预训练方法相比,UPNN对不同类型的成像对象和硬件更具灵活性。此外,实验结果和模拟分析证明,UPNN在图像质量、横向分辨率和稳健性方面优于传统的迭代叠层成像算法。