Holographic imaging has long grappled with the complexities of dynamic environments, where unpredictable distortions pose significant obstacles. Traditional deep learning methods, which heavily rely on specific data conditions, often struggle to adapt in such diverse scenes. Recognizing this limitation, researchers at Zhejiang University embarked on a pioneering exploration at the intersection of optics and deep learning. Their objective was to uncover the indispensable role of physical priors in ensuring the alignment between data and pre-trained models.
The researchers delved into the impact of spatial coherence and turbulence on holographic imaging. Spatial coherence refers to the orderly behavior of light waves. In chaotic conditions, such as those found in turbulent environments, holographic images become blurry and noisy, carrying less information. Maintaining spatial coherence is therefore essential for clear and high-quality holographic imaging.
To address the challenges presented by dynamic environments, the Zhejiang University researchers developed an innovative method called TWC-Swin, which stands for “train-with-coherence swin transformer.” This method leverages spatial coherence as a physical prior to guide the training of a deep neural network. The network, built on the Swin transformer architecture, excels at capturing both local and global image features.
To validate the effectiveness of their approach, the researchers designed a light processing system that produced holographic images with varying degrees of spatial coherence and turbulence. These holograms, created from natural objects, were used as training and testing data for the neural network. The results of the experiments demonstrated that TWC-Swin restores holographic images with exceptional quality, even under conditions of low spatial coherence and arbitrary turbulence. Importantly, the method exhibited strong generalization capabilities, extending its application to previously unseen scenes not included in the training data.
This groundbreaking research not only addresses image degradation in holographic imaging across diverse scenes but also signifies a remarkable fusion of physical principles and deep learning. By integrating optics and computer science, the study sheds new light on the potential for a successful synergy between these fields. As the research unfolds, it opens up exciting possibilities for enhancing holographic imaging, empowering us to perceive through the turbulence with unprecedented clarity.