Super-resolution imaging of terahertz rotating coherent scattering microscopy based on deep-learning method

Terahertz Rotating Coherent Scattering (ROCS) method can realize the super-resolution imaging through the illumination of the evanescent wave. Under one illumination direction, it is treated as a coherent imaging process, and the local destructive interference may improve the imaging resolution. Then, the intensity images under different illumination directions are obtained by rotating the sample for several times, and they are added up incoherently to achieve the final result. However, in the experiment, the operation of rotating the sample is inconvenient and the rotational angle deviation, unstable light source output may exist to affect the results. In addition, simply summing limited intensity images may cause loss of sample information. To solve the above problems, we combine the deep learning method with the ROCS to realize the super-resolution imaging. Firstly, the reconstruction result is compared between the single-input and the multi-input network models, and the simulation results show that the latter one has the better quality. Then, with regard to that there is usually not enough experimental data, the transfer learning technique is introduced to the deep learning method. Use a large amount of source data and a small amount of target data to complete the training of the network. The simulation results show that the quality of reconstruction image is significantly improved. It implies that the ROCS imaging combining with deep learning method can further improve the resolution and the contrast of the reconstruction result.