Lensfree auto-focusing imaging with coarse-to-fine tuning method

Sample-to-sensor distance is a crucial predefined parameter for lensfree on-chip microscopes. Inaccurate estimation of the distance could lead to a failure of high-frequency detail recovery. To address it, conventional methods either evaluate the sharpness of back-propagated Z-stack holograms or adopt a pretrained end-to-end classification neural network. However, fast and robust distance estimation is still a challenging task. Here we propose a coarse-to-fine auto-focusing method to achieve fast and robust distance estimation for lensfree on-chip microscopy. In our method, a simulation-driven focus network (sFocusNet) is designed as a coarse tuning to decrease the distance searching range, and then a regularization of gradient (RoG) metric is constructed as a fine tuning to achieve an accurate estimation. Experimental results of different samples are given to verify the superiority and generalization of our method. We believe that our method will offer a practical auto-focusing solution for the commercialization of lensfree microscopes.