Propagation as Data (PaD) : neural phase hologram generation with variable distance support

다국어 초록 (Multilingual Abstract)

Recent deep-learning-based methods for generating phase holograms have seen significant advancements. However, most of the neural network models developed so far are trained and validated only for a single set of propagation parameters such as a distance, a wavelength and a pixel pitch. Consequently, if any of these parameter values are altered, the performance of these models tends to decline significantly unless re-trained with the new parameters. To address this, we introduce a novel approach called ’Propagation as Data (PaD)’ that can generate high-quality phase holograms across a range of parameter sets with a single trained model. Unlike conventional methods, our model does not include the wave propagation process inside the model. Instead, we pre-calculate propagation kernels for a range of parameter sets and use them as additional inputs to our model.
Our PaD model consists of an pre-trained Image Encoder, a set of shallow Kernel Encoders, and a Phase Decoder. The Image Encoder extracts features from the object, while the Kernel Encoders extract features from the pre-calculated propagation kernels. These object and kernel features are then fused together in the frequency domain using the Fast Fourier Convolution (FFC) within the Phase Decoder. By using the propagation kernels as a conditioning data, our model could easily adapt to the phase pattern of the object under different propagation settings, enabling high-quality phase hologram generation with a single model on-the-fly. Experimental results show that our PaD model significantly improves reconstruction quality against existing models when tested across multiple propagation distances. More specifically, 5.596dB and 5.489dB average PSNR improvement are observed against HoloNet and DPRC, respectively.

목차 (Table of Contents)

• 1 Introduction 1
• 2 Related Works 5
• 2.1 Wave Propagation 5
• 2.2 Phase-only Hologram Generation 6
• 3 Propagation as Data (PaD) 7
• 3.1 Pre-trained Image Encoder 9
• 3.2 Kernel Encoder & Phase Decoder 10
• 3.3 Implementation, Training & Inference details 11
• 4 Experimental Results 13
• 4.1 Experimental Settings 13
• 4.2 Results of Full-color Variable Distance Generation 14
• 4.3 Comparison with Per-channel Single Distance Generation 16
• 4.4 Ablation Studids 18
• 5 Conclusion 21
• Reference 22