PDAF is a technique that estimates the in-focus position according to the in-put pair of phase images. Most previous PDAF methods are noise sensitive, so noisyphase information usually lead to slow or inaccurate autofocus. To address this is-sue, I designed a model, called AF-Net, based on convolutional neural network (CNN).The AF-Net accurately estimates the correspondence between phase images and es-timates the in-focus position accordingly. In most cases, the AF-Net reaches the in-focus position in two lens movements regardless of the noise. Our work on PDAF has been published in IEEE Transactions on Image Processing and Electrical Imaging (see publication). Our demo videos of this work are available in the following links:
Demo: AF-Net v.s. iPhoneSE
Demo: AF-Net v.s. iPhone7

Awards

First Prize in NTU Undergraduate Innovation Award
College Student Research Creativity Award