Physics-based generative adversarial network for real-time acoustic holography

Abstract

Acoustic holography (AH) encodes the acoustic fields in high dimensions into two-dimensional holograms without information loss. Phase-only holography (POH) modulates only the phase profiles of the encoded hologram, establishing its superiority over alternative modulation schedules due to its information volume and storage efficiency. Moreover, POH implemented by a phased array of transducers (PAT) facilitates active and dynamic manipulation by independently modulating the phase of each transducer. However, existing algorithms for POH calculation suffer from a deficiency in terms of high fidelity and good real-time performance. Thus, a deep learning algorithm reinforced by the physical model, i.e. Angular Spectrum Method (ASM), is proposed to learn the inverse physical mapping from the target field to the source POH. This method comprises a generative adversarial network (GAN) evaluated by soft label, which is referred to as soft-GAN. Furthermore, to avoid the intrinsic limitation of neural networks on high-frequency features, a Y-Net structure is developed with two decoder branches in frequency and spatial domain, respectively. The proposed method achieves the reconstruction performance with a state-of-the-art (SOTA) Peak Signal-to-Noise Ratio (PSNR) of 24.05 dB. Experiment results demonstrated that the POH calculated by the proposed method enables accurate and real-time hologram reconstruction, showing enormous potential for applications.