|
Computer Generated Holography (CGH) promises unprecedented capabilities for a variety of applications in Optics and Photonics. However, one of the biggest challenges for CGHs is the fundamental tradeoff between algorithm runtime and achieved reconstruction fidelity and efficiency while maintaining light projections at real-time frame rates. In addition, the light projection quality achieved by most CGH-modalities are rather low due to the mismatch between the optical wave propagation of the applied Spatial Light Modulator (SLM) and its simulated model. A promising new avenue of CGH, neural holography, utilizes machine learning models in the generation of single and multi plane holograms. Neural network generated holograms have the distinct advantage that inference is performed in constant time without the need for iterative calculations of the phase SLM pattern. This allows the networks to generate holograms 3-500 times faster than traditional iterative algorithms, which enables the applications dependent on real-time holography. State-of-the-art implementations of neural holography [1, 2] furthermore achieve higher accuracy than traditional iterative algorithm, when compared to target images. Applications of these SLM-encoded CGHs include all areas where a fast and parallel one- or two-photon light excitation is needed such as in Laser Material Processing, Additive Manufacturing and 3D printing, Neurophotonics and Optogenetics, Laser Image Projection and many more.
|
