Automatic Lens Design based on Differentiable Ray-tracing.

Xinge Yang, Qiang Fu, Wolfgang Heidrich
OSA Imaging and Applid Optics Congress - Computational Optical Sensing and Imaging (COSI), 2022.



Pipeline of our proposed differentiable ray-tracing model. Starting from a randomly generated design, our model can optimize lens parameters and positions for the best imaging quality. Attention window is dynamically adjusted for a faster training speed and getting out from local minimums.

Abstract

We propose a fully differentiable optical design method enabled by curriculum learning. Preliminary results that our framework is suitable to solve highly non-convex problems like cellphone lens design.

Paper and Video

paper          [Yang2022AutoLens.pdf] 
code           [https://github.com/vccimaging/AutoLens]