THz rotational coherent scattering super-resolution imaging based on deep convolutional neural networks

Terahertz (THz) has a wide range of applications in imaging due to its penetrating properties, low photon energy, and extreme absorbability by water molecules. Its practical application is limited by its inherently long wavelength, relatively few modulation devices, and noisy imaging systems. THz-based rotationally coherent scattering (ROCS) imaging is a newly developed super-resolution imaging method, which uses a high-resistance silicon hemispherical lens to generate the evanescent wave to illuminate the sample. However, many intensity images acquisition through some kind of the rotation are necessary for achieving high-quality imaging, and thus it becomes time-cost and prone to the degraded results because of angle errors, displacements. In the same time, the unstable output of the THz light source will also negatively affect the results. In this work, we propose a convolutional neural networkbased super-resolution imaging method for THz ROCS imaging. The multi-input network is constructed, and only six low-resolution images with different azimuthal illumination are needed to achieve reconstruction results that surpasses conventional THz ROCS imaging. Since it is not possible to directly acquire a large number of highand low-resolution image pairs to train the network, we combine transfer learning with computer-generated 115 sets of square periodic metal gratings and 5 sets of circular aperture periodic metal gratings to complete the training of the Unet. The simulation results show that by freezing the encoder of the pre-trained network model, the decoder is randomly initialized and re-trained, and the network reconstructs the image with specifically sharper edges, fewer artifacts, and more detailed information. Using only six low-resolution intensity images with different azimuthal illumination, we achieve super-resolution reconstruction of a $70 \mu \mathrm{~m}$ circular aperture periodic metal grating. The reconstructed image quality superior to conventional THz ROCS imaging results.