Near-infrared high-resolution imaging via deep learning based on a broadband achromatic metalens

Abstract

The development of large-aperture broadband achromatic metalenses is crucial for advancing compact optical systems, but it remains challenging due to the complex phase compensation requirements and the inherent trade-off between numerical aperture and aperture size. Therefore, a high-resolution near-infrared imaging framework that integrates a large-aperture polarization-independent achromatic metalens with a deep learning-based super-resolution algorithm is proposed. The metalens, operating from 1.34 to 1.54 µm with a diameter of 210 µm, is designed using a discrete multi-wavelength methodology combined with a global optimization algorithm, thereby breaking the limitations of conventional continuous achromatic design. Experimental results show that the fabricated metalens achieves a resolution of 19.69 µm. To further transcend the hardware limitation, a super-resolution residual network (SRResNet) is employed for post-processing, enhancing the imaging resolution from 19.69 µm to 15.63 µm. This work presents a feasible scheme by combining meta-optics with deep learning, which exhibits promising potential in applications such as virtual/augmented reality and integrated photonics.