Publication Details
Overview
 
 
Makoto Sekiguchi, Chung-Hsuan Huang, David Blinder, Shosei Yabuki, Fan Wang, Peter Schelkens, Han-Yen Tu, Chau-Jern Cheng, Tomoyoshi Ito, Tomoyoshi Shimobaba
 

Contribution to journal

Abstract 

Recent advances in computer-generated holography (CGH) have significantly improved visual quality through high-resolution rendering; however, the accompanying increase in data size has become a critical obstacle to practical deployment. Conventional image compression techniques such as JPEG and high efficiency video coding (HEVC) do not adequately account for the unique statistical and spectral characteristics of holograms, thereby limiting both compression efficiency and reconstruction quality. This study proposes an efficient compression framework specialized for digital holograms by integrating a ringing reduction technique for diffraction calculations with CompressAI, a deep learning-based image compression framework. We demonstrate that ringing artifacts are a major factor hindering efficient hologram compression, and show that neural networks can achieve superior performance by learning the intrinsic statistical and spectral features of holograms. Our method achieves a superior balance between compression efficiency and reconstruction quality compared to conventional approaches, particularly at low bit rates. Furthermore, by introducing a multi-channel input representation, our method achieves higher compression ratios.

Reference