Res-U2Net: untrained deep learning for phase retrieval and image reconstruction.

Area of Science:

• Optics and Photonics
• Computer Vision
• 3D Imaging

Background:

• Deep learning for image reconstruction typically requires large datasets, which are often impractical to acquire.
• Untrained deep learning methods offer a solution by learning to invert the physical image formation process without prior data.
• Phase retrieval is crucial for extracting detailed object information, but conventional methods face data limitations.

Purpose of the Study:

• To introduce a novel untrained deep learning model, Res-U2Net, for phase retrieval applications.
• To leverage extracted phase information for precise surface change detection and 3D mesh generation.
• To evaluate the performance of the Res-U2Net model against existing UNet and U2Net architectures.

Main Methods:

• Development of an untrained Res-U2Net deep learning architecture for phase retrieval.
• Utilizing the phase information derived from the model to analyze object surface variations.
• Generating a mesh representation to reconstruct the 3D structure of the object.
• Comparative analysis using the GDXRAY dataset, benchmarking against UNet and U2Net models.

Main Results:

• The Res-U2Net model demonstrates effective phase retrieval without requiring a large training dataset.
• Extracted phase information accurately reflects changes in object surfaces.
• Successful generation of 3D mesh representations from the retrieved phase data.
• Performance comparison indicates the efficacy of Res-U2Net in phase retrieval tasks.

Conclusions:

• The proposed untrained Res-U2Net model is a viable and data-efficient solution for phase retrieval.
• This method facilitates accurate 3D surface reconstruction and analysis.
• The findings suggest potential for broader applications in scientific imaging and metrology where data is scarce.