Deep-learning-enhanced Digital Holographic Autofocus Imaging

Digital holography can achieve automatic focusing and quantitative imaging in the whole field of view, and has been widely used in imaging measurement tasks. Conventional digital holographic autofocusing algorithms are usually iterative approaches with time-consuming computational process. In recent years, deep learning technology has been applied in digital holography. However, most current findings about this field are deep learning algorithms dealing with partial operations of digital holographic reconstruction, where angular spectrum propagation and background images are still needed in reconstruction. Inspired by U-Net and residual network (ResNet), a new convolutional neural network (CNN) is proposed in this paper to realize digital holographic autofocus imaging. After proper training, the proposed CNN can obtain the focused reconstructed results by performing a feed forward propagation, and no background images are needed in reconstruction process.