Point spread function-inspired deformable convolutional network for holographic displays

Convolutional neural networks (CNNs) have found extensive application in computer-generated holography (CGH). Nonetheless, CNNs possess limited capability to effectively model intricate geometric transformations between object points and their corresponding point spread functions due to the constrained structures of fixed convolutional kernels. In order to address this issue, we propose deformable holography (DeH) algorithm for CGH. We demonstrate that utilizing deformable convolutions enable adaptive modeling of geometric transformations. The proposed DeH algorithm generates high-quality 1080P 3D holograms in real-time, consistently outperforming existing approaches. We also validate our approach on an experimental prototype holographic display, and demonstrate DeH algorithm's ability to accurately reconstruct 3D scenes. Overall, our work introduces new possibilities of utilizing deformable convolutions for deep learning in the realm of holographic displays.