Cycle-consistent physical constraint deep learning for incoherent self-interference digital holographic reconstruction

Abstract

We propose a novel incoherent holographic adaptive reconstruction network. This network employs a dual-input fully connected architecture for adaptive feature optimization and introduces an innovative iterative optimization framework to address physical consistency constraints. By using cyclic consistency loss with nonlinear deconvolution embedding, our method achieves a significant improvement over traditional reconstruction methods, realizing a 41.42% enhancement in lateral resolution while increasing the peak signal-to-noise ratio of the reconstructed images by 3.2 times. Comprehensive evaluations conducted on the DIV2K dataset demonstrate that our approach maintains exceptional reconstruction quality compared to existing supervised learning models. This method enables high-quality single-shot holographic reconstruction under data-scarce conditions, offering potential applications for real-time biological imaging and industrial inspection.