Low-Cost Acoustic Field Reconstruction With Physics-Incorporated Deep Learning for Binary Amplitude-Only Hologram

Qing Liu, Chengxi Zhong, Zhenhuan Sun, Youfu Li, Hu Su, Song Liu

Abstract: Acoustic reconstruction aims to recreate target acoustic fields by spatially modulating acoustic waves and has significant application potential. On basis of acoustic holography (AH), this article focuses on low-cost acoustic reconstruction technique using binary amplitude-only hologram (BAOH). A novel physics-incorporated deep learning framework trained with a two-stage strategy is proposed, achieving outstanding accuracy and real-time performance in BAOH prediction. Specifically, we introduce the binary U-net (BU-net) architecture, which combines the classical U-net with a customized binary layer. With the unique design, BU-net is capable of yielding binary results without being hindered by the gradient invalidation. By integrating the acoustic wave propagation model, BU-net is trained to learn the inverse mapping from the target acoustic field to the corresponding source BAOH, also eliminating the need for labor-intensive annotation collection. The simulation experiments show that the proposed method can reduce the influence of quality degradation from binarized source holograms and achieved satisfactory reconstruction quality. Comparison of the experiments further demonstrate the superiority of our proposed method over state-of-the-art (SOTA) method in the aspects of both accuracy and real-time performance. Finally, physical experiments confirm the alignment with simulation results, showcasing promising potential for real-world applications.