Opt Lett. 2021 Jul 1;46(13):3123-3126. doi: 10.1364/OL.422684.
Terahertz (THz) imaging has been applied successfully in numerous applications, from medical imaging to industrial non-destructive detection. However, low resolution has always been a problem due to its long wavelength. A convolution neural network (CNN) is quite effective at improving the resolution of images in optics, in which real numbers are manipulated corresponding to measured intensity. Compared to optics, it is quite feasible to gain both the amplitude and phase information in THz imaging. In this Letter, we have extended the CNN from a real number domain to a complex number domain based on the wave nature of THz light. To the best of our knowledge, this is the first time that such a complex convolution neural network (CCNN) has been shown to be successful in THz imaging. We have proved that resolution can be 0.4 times of the beam size via this approach, and half a wavelength resolution can be obtained easily. Compared to the CNN, the CCNN generates an extra 27.8% increase in terms of contrast, implying a better image. Phase information can be recovered well, which is impossible for the CNN. Although the network is trained by the MNIST dataset, it is quite powerful for image reconstruction. Again, the CCNN outperforms the CNN in terms of generalization capability. We believe such an approach can help to overcome the lower-resolution bottleneck in THz imaging, and it can release the requirement of critical optical components and extensive fine-tuning in systems. THz biomedical imaging, non-destructive testing (NDT), and a lot of imaging applications can benefit from this approach.
太赫兹（THz）成像已经成功应用于许多领域，从医学成像到工业无损检测。然而，由于其波长较长，分辨率低一直是一个问题。卷积神经网络（CNN）在提高光学图像分辨率方面非常有效，在光学中，实数对应于测量强度。与光学相比，在太赫兹成像中同时获得幅度和相位信息是非常可行的。在本信中，我们基于太赫兹光的波动性，将 CNN 从实数域扩展到复数域。据我们所知，这是首次成功将这种复卷积神经网络（CCNN）应用于太赫兹成像。我们已经证明，通过这种方法可以将分辨率提高到光束尺寸的 0.4 倍，并且很容易获得半波长分辨率。与 CNN 相比，CCNN 的对比度提高了 27.8%，这意味着图像质量更好。可以很好地恢复相位信息，这是 CNN 无法做到的。虽然该网络是通过 MNIST 数据集进行训练的，但它在图像重建方面非常强大。同样，CCNN 在泛化能力方面优于 CNN。我们相信这种方法可以帮助克服太赫兹成像中的低分辨率瓶颈，并释放对系统中关键光学元件和广泛微调的要求。太赫兹生物医学成像、无损检测（NDT）和许多成像应用都将受益于这种方法。
