Two-step training deep learning framework for computational imaging without physics priors.

Area of Science:

• Computational imaging
• Deep learning applications
• Image reconstruction

Background:

• Deep learning (DL) is vital in computational imaging, often using preprocessors with physics priors.
• Model mismatches and network over-parameterization pose challenges in DL-based image reconstruction.
• Ill-posed inverse problems in imaging can lead to suboptimal DL performance.

Purpose of the Study:

• To propose a two-step training deep learning (TST-DL) framework for computational imaging that bypasses the need for physics priors.
• To address challenges of model mismatch and network over-parameterization in image reconstruction.
• To develop a flexible and robust DL approach for diverse computational imaging systems.

Main Methods:

• A two-step training DL (TST-DL) framework was developed.
• The first step involves training a fully-connected layer (FCL) to learn the inverse model directly from raw measurement data.
• The second step concatenates the pre-trained FCL with an untrained U-Net architecture for image optimization.

Main Results:

• The TST-DL framework demonstrated performance comparable to methods using perfect physics knowledge.
• It showed robustness to noise, model ill-posedness, and model mismatch.
• TST-DL outperformed end-to-end training with less overfitting and better results than approaches with imperfect physics knowledge.

Conclusions:

• The TST-DL framework offers a flexible approach to image reconstruction without requiring explicit physics priors.
• This method effectively mitigates issues related to model mismatch and network over-parameterization.
• TST-DL is a promising technique for various computational imaging applications.