Simplified design method for optical imaging systems based on deep learning

Modern optical design methods pursue achieving zero aberrations in optical imaging systems by adding lenses, which also leads to increased structural complexity of imaging systems. For given optical imaging systems, directly reducing the number of lenses would result in a decrease in design degrees of freedom. Even if the simplified imaging system can satisfy the basic first-order imaging parameters, it lacks sufficient design degrees of freedom to constrain aberrations to maintain the clear imaging quality. Therefore, in order to address the issue of image quality defects in the simplified imaging system, with support of computational imaging technology, we proposed a simplified spherical optical imaging system design method. The method adopts an optical-algorithm joint design strategy to design a simplified optical system to correct partial aberrations and combines a reconstruction algorithm based on the ResUNet++ network to correct residual aberrations, achieving mutual compensation correction of aberrations between the optical system and the algorithm. We validated our method on a two-lens optical imaging system and compared the imaging performance with that of a three-lens optical imaging system with similar first-order imaging parameters. The imaging results show that the quality of reconstructed images of the two-lens imaging system has improved (SSIM improved 13.94%, PSNR improved 21.28%), and the quality of the reconstructed image is close to the quality of the direct imaging results of the three-lens optical imaging system.