Can Deep Learning Revolutionize Phase Recovery? This Review Paper Explores Its Impact and Future in Computational Imaging

Light is studied in two essential components:  amplitude and phase. However, optical detectors that rely on photon-to-electron conversion face problems capturing the phase due to their restricted sampling frequency. The limitation they face is that while they can easily measure the amplitude, they struggle to grasp the phase due to limitations in their sampling frequency. However, this can be problematic because the phase of the light field contains important information. Therefore, accurately recovering the phase of the light field is vital for determining the structure of the samples.

Researchers earlier used to use several traditional methods for phase recovery. These methods include holography/interferometry, Shack-Hartmann wavefront sensing, transport of intensity equation, and optimization-based methods. These methods, though useful, had several problems in each technique, such as low spatiotemporal resolution and high computational complexity. 

Consequently, researchers of The University of Hong Kong, Northwestern Polytechnical University, The Chinese University of Hong Kong, Guangdong University of Technology, and Massachusetts Institute of Technology in a recent review paper published in Light: Science & Applications reviewed using deep learning for phase recovery from four perspectives. The first perspective discussed using deep learning to pre-process intensity measurements before phase recovery. Some of the pre-processing techniques include pixel super-resolution, noise reduction, hologram generation, and autofocusing. These techniques help improve the quality of the input data and can improve phase recovery results.

In the second perspective, the researchers focused on the Deep-learning-post-processing technique for phase recovery. They used deep learning during the phase recovery process. The neural networks perform phase recovery independently or alongside a physical model in this method. This approach has the benefit of providing faster and more accurate phase recovery than traditional methods. The third perspective is deep learning for post-processing after phase recovery. It has noise reduction, resolution enhancement, aberration correction, and phase unwrapping techniques. These techniques can improve the accuracy of the recovered phase. Finally, the fourth perspective explores using the recovered phase for specific applications, such as segmentation, classification, and imaging modality transformation. This approach helps to get valuable insights from the recovered phase into the properties and behavior of the sample under investigation.

The researchers emphasize that while using this deep learning technique for this task has numerous benefits, it has certain limitations, too, as it also has certain risks. They highlight that while some methods may appear similar, they have subtle differences that are challenging to detect. They suggest combining physical models with deep neural networks to overcome these risks, particularly when the physical model closely aligns with reality. This increases the overall accuracy of the method.

In conclusion, this technique of using deep learning for phase recovery has significant advantages over traditional phase recovery methods as it has enhanced speed, accuracy, and versatility. As researchers try to improve the technique, the system’s limitations will also be solved. By doing so, researchers can unlock the potential of deep learning for phase recovery and advance the understanding of complex systems in diverse fields.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our 35k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.