Artifact-Resilient Real-Time Holography

Obstructions in Random vs. Smooth Phase Holography

Recent camera-in-the-loop (CITL) optimization has enabled holographic displays to correct for static aberrations like dust and imperfect lenses, producing high-quality 3D images. However, this approach fails when faced with dynamic obstructions such as eyelashes, which can completely block the view. This limitation highlights a fundamental trade-off in hologram generation between using a smooth phase pattern versus a more random-like phase. Historically, smooth phase has been preferred for its ability to create images with high contrast and minimal speckle. The critical drawback is that it concentrates light into a very small viewing window, or “eyebox,” making the hologram fragile and easily blocked. A more random phase, in contrast, distributes light over a wider angle, allowing the image to diffract around unanticipated obstacles.

Introducing: Rayleigh Distance for Quantifying Phase Randomness

Although random phase can naturally lead to artifact-resilient holograms, this notion becomes ill-defined in settings such as neural network based phased CGH, where a clear “phase initialization” does not exist. To address this, we introduce a metric for phase randomness that performs patch-wise similarity between a hologram’s eyebox intensity and that of a reference random-phase hologram whose eyebox intensity follow a Rayleigh distribution. We denote this loss by $\mathcal{L}_{RD}$ ​ (Rayleigh Distance). Below, we contrast a constant-phase hologram (left) with several optimized variants: (i) constant phase with MSE loss only; (ii) constant phase with MSE plus a uniform-intensity eyebox loss, $\mathcal{L}_{UD}$; (iii) constant phase with $\mathcal{L}_{RD}$ ​ ; and, as an upper reference, (iv) an unoptimized hologram initialized with uniformly random phase (right). Empirically, $\mathcal{L}_{RD}$ ​ steers optimization toward solutions exhibiting higher phase randomness and a broader energy distribution in the eyebox, even when starting from constant phase. Moreover, we observe that full randomness is not required for artifact resilience: partially randomized solutions (second from right) already provide substantial protection.

Real-Time Artifact-resilient Holography

Using $\mathcal{L}_{RD}$, we are able to train a neural network based CGH phase retrieval engine that outputs artifact-resilient holograms in real time. We call our method Real-Time ARH. We compare HoloNet vs. our Real-Time ARH with a quantized phase light modulator (TI DLP6750Q1EVM). Without obstructions, our Real-Time ARH produces better visual quality, as HoloNet requires camera-in-the-loop calibration to overcome the hardware imperfections of the real setup. Our method, because it relies on random phase holography, is naturally more robust to these imperfections. When obstructions like eyelashes are introduced, HoloNet holograms are distorted. Our Real-Time ARH, on the other hand, is artifact-resilient, since its use of random phase holography allows light to scatter around obstructions. These experimental captures use 8-frame time multiplexing per channel.

Training Artifact-Resilient Hologram Networks