Terahertz deep-optics imaging enabled by perfect lens-initialized optical and electronic neural networks

Terahertz imaging has shown great potential in various fields and applications, such as non-destructive testing, security screening and biomedical researches. However, many factors, such as diffraction effects and various geometric aberrations, have severely limited the performance of refractive imaging systems in terahertz range, bringing in noise, blur and distortion into the captured images. In this work, we propose a novel intelligent optical modulator, named as perfect lens-initialized optical neural network (PLIONN), to facilitate the high-quality terahertz imaging. By combining the phase profile of conventional refractive lens with a trainable optical model, the PLIONN model is able to incorporate both the powerful imaging capability of traditional refractive lens and the data-driven iterative optimization into the design of imaging lens, highly promoting the improvement of spatial resolution and imaging quality. Moreover, a simple electronic neural network (ENN) is adopted to further deal with the image degradation computationally, and correspondingly, a stage-united training scheme is proposed to connect the optical imaging with the post-processing. Therefore, the proposed opto-electronic framework constitutes a dual-core setup by highlighting the intelligent computing in both optical and electronic system, which is more robust and flexible than the existing single-core counterpart. Simulation on both Siemens star resolution chart and various imaging targets have demonstrated the superiority of such framework.