Chinese Scientists Develop New Deep Learning Imaging Technology

Researchers at the Xi'an Institute of Optics and Precision Mechanics (XIOPM) under the Chinese Academy of Sciences have made significant progress in developing an interpretable deep learning method for reconstructing high-resolution images from Fourier ptychographic imaging. Fourier ptychographic imaging is an emerging computational technique that combines low-pass filtering of the pupil function, scanning and sampling in the frequency domain, Fourier transformation, and sophisticated noise contamination. Traditional approaches using deep neural networks, such as convolutional neural networks (CNNs), struggle with the increased complexity of environmental noise in long-distance scenarios, making high-resolution image reconstruction particularly challenging. To address this issue, the research team at XIOPM introduced an optimization model that integrates learnable regularization terms. This model, combined with a proximal gradient optimization algorithm, provides a strong theoretical foundation for building interpretable deep learning models. Building on this, they designed the Model-Driven Fourier Ptychography Network (MDFP-Net). MDFP-Net is the first deep unfolding network to alternate optimization between the complex and real domains, effectively embedding the amplitude flow reconstruction algorithm into its architecture. This novel design allows for the successful reconstruction of amplitudes with clear physical significance, enhancing both the theoretical soundness and performance of deep learning in computational imaging. To validate the effectiveness of MDFP-Net in achieving high-quality and rapid image reconstruction, the team developed a long-distance reflective Fourier ptychographic imaging system. They successfully captured real sample data from a distance of 8.7 meters. The results not only deepen the understanding of Fourier ptychographic imaging techniques but also introduce a new technical approach by merging deep learning with computational imaging methods. This breakthrough has significant implications for the field of computational optical imaging. The paper detailing their findings, titled "MDFP-Net: A Model-Driven Deep Neural Network for Fourier Ptychography," has been published in the journal Computational Visual Media. The research was supported by the CAS Western Young Scholars Program and the Open Fund of the National Key Laboratory of Ultrafast Optical Science and Technology. Key features of MDFP-Net include its modular design, where each component has a clear physical meaning, and the ability to handle the increased noise complexity encountered in long-distance imaging. The team's experimental setup and results demonstrate that MDFP-Net can reconstruct high-resolution images efficiently, even under challenging conditions. By providing a more interpretable and robust framework for Fourier ptychographic imaging, MDFP-Net opens up new possibilities for advancing computational optical imaging technology.