Unwrapping of phase-only holographic data using a convolutional neural network

In this work, we introduce a modified convolutional neural network (CNN) based on a U-Net, capable of unwrapping the phase of both computer-generated holograms (CGHs) and digital holograms of a broad range of objects. We introduce a structural similarity index measurement (SSIM) based loss function and a Gaussian activation function to modify a CNN for phase unwrapping. We train our network with wrapped-unwrapped phase pairs obtained through an iterative Fresnel CGH algorithm and a quality guided path (QGP) unwrapping method. After training, we input the wrapped phase of any CGH into the CNN and obtain an unwrapped phase that allows for successful reconstruction of the original object with limited degradation. Furthermore, our CNN can perform unwrapping faster compared to the direct application of the QGP algorithm. We also demonstrate that the CNN has excellent generality, with capability for successful unwrapping of CGH corresponding to different object-hologram plane distances, different types of objects, and even experimentally recorded Fourier and Fresnel digital holograms.