Reconstruction of partially obscured objects with a physics-driven self-training neural network

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Imaging Science

Background:

• Terahertz (THz) imaging offers unique capabilities for non-ionizing 3D scene reconstruction.
• Traditional holographic imaging requires phase information, which is often unavailable.
• Reconstructing 3D scenes from power-only THz data presents significant challenges due to missing phase information.

Purpose of the Study:

• To develop an artificial intelligence-supported method for 3D holographic imaging using coherent terahertz radiation.
• To reconstruct 3D scenes from intensity-only measurements, overcoming the limitation of phase retrieval.
• To address challenges in reconstructing partially obscured objects and minimizing interference in THz imaging.

Main Methods:

• A physics-informed deep learning (DL) algorithm was developed, incorporating angular spectrum theory as prior knowledge.
• A synthetic dataset of diffraction patterns was generated using object information.
• A physics-informed neural network (NN) was self-trained using both synthetic and unlabeled experimental THz data.
• A recursive training strategy was employed, where the NN iteratively predicted and reincorporated images into the training set.

Main Results:

• The DL approach successfully reconstructed 3D scenes from power-only THz images.
• The method demonstrated effectiveness in reconstructing partially obscured objects.
• Mutual interference during object reconstruction was minimized.
• The approach showed robustness and effectiveness in data-scarce scenarios.

Conclusions:

• The proposed physics-informed DL method significantly advances 3D THz imaging capabilities.
• This technique provides a viable solution for reconstructing 3D scenes when phase information is absent.
• The self-training strategy enables effective learning from limited and unlabeled experimental data, paving the way for practical applications.