Restoration of infrared metalens images with deep learning

Abstract

Aberrations and noise hinder the efficient use of infrared metalenses. We have developed a method to improve infrared metalens-produced images by incorporating transfer learning with a cycle generative adversarial network (CycleGAN). By simulating metalens-induced aberrations on images from the ImageNet dataset, we created a pre-training dataset for transfer learning, which improves the capabilities of the CycleGAN model. Our method enhances the peak signal-to-noise ratio (PSNR) and contrast of the images by an average of 5.07 dB and 63.81, respectively, in the spatial domain compared to the original images captured using the infrared metalens. Furthermore, the introduction of transfer learning improves the erratic translation and the model's ability to restore missing details, and enhances the edge intensity and similarity in the frequency domain by 8.85 % and 3.19 %, respectively.