Graph-enhanced global dependency modeling for complex wavefront retrieval.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computational Imaging

Background:

• Convolutional neural networks (CNNs) are standard for complex wavefront retrieval (CWR).
• CNNs' local receptive fields limit their ability to model global dependencies in wavefront data, causing performance issues.
• Existing methods struggle with capturing long-range correlations crucial for accurate wavefront reconstruction.

Purpose of the Study:

• To introduce a novel graph-enhanced neural network (GENet) for improved CWR.
• To address the limitations of CNNs in modeling global wavefront dependencies.
• To enhance the representational capacity of neural networks for wavefront data.

Main Methods:

• Integrating graph convolution with a CNN backbone to create GENet.
• Employing a topology-aware design that leverages the physical characteristics of wavefront data.
• Enabling message passing across non-local, semantically relevant regions for enhanced feature extraction.

Main Results:

• GENet significantly outperforms traditional GS (Gerchberg-Saxton) and CNN-based methods.
• Achieved an amplitude SSIM of 0.78 and phase RMSE of 0.52 radians in a representative test case.
• Demonstrated superior amplitude and phase reconstruction accuracy compared to GS (0.22/1.15) and IPMnet (0.42/1.08).

Conclusions:

• GENet offers a robust and high-precision solution for complex wavefront retrieval.
• The graph-enhanced approach effectively models global dependencies in wavefront data.
• GENet shows significant potential for practical applications requiring accurate wavefront reconstruction.