[9p-PB6-1~21] 3.2 Information photonics and image engineering

[9p-PB6-18] Physics-Driven Phase Optimization for Holograms Using U-Net with Bottleneck KAN

〇Yuito Watanabe¹, Kai Kumano¹, Fan Wang¹, Tomoyoshi Ito¹, Tomoyoshi Shimobaba¹ (1.Chiba Univ.)

Keywords:

hologram,KAN,deep-learning

To tackle the image quality degradation in hologram phase imaging, we propose a new physics-driven learning framework that introduces KAN, which has excellent nonlinear representation capability, into the bottleneck layer of U-Net and combines it with our original SSIM loss function. In tests using CIFAR-10, the KAN based on Morlet Wavelet showed higher PSNR and SSIM compared to the conventional U-Net, proving its advantage in improving image quality. We also extend this method to 3D objects.