On the interplay between physical and content priors in deep learning for computational imaging

Area of Science:

• Computational Imaging
• Deep Learning
• Phase Retrieval

Background:

• Deep learning (DL) excels in computational imaging but generalization and interpretability remain challenges.
• Understanding if DL models learn physics or merely memorize data is crucial for reliable applications.
• The Phase Extraction Neural Network (PhENN) is a deep neural network (DNN) used for quantitative phase retrieval.

Purpose of the Study:

• To investigate the relationship between training data characteristics and DL model generalization and interpretability.
• To explore how training data influences the learning of underlying physics models in phase retrieval.
• To connect training data entropy to regularization effects and model performance.

Main Methods:

• Utilized the PhENN model for quantitative phase retrieval in lensless imaging.
• Analyzed the impact of training dataset properties, specifically Shannon entropy, on model behavior.
• Correlated training data entropy with regularization strength and the learning of physical models.

Main Results:

• The choice of training examples is critical for DL generalization and interpretability.
• Higher training image entropy leads to weaker regularization effects.
• Weaker regularization allows for better learning of the underlying propagation model (weak object transfer function).
• DNNs trained on high-entropy datasets (e.g., ImageNet) show superior cross-domain generalization compared to those trained on low-entropy datasets (e.g., MNIST).

Conclusions:

• Training data entropy directly impacts DL model generalization and its ability to learn physical principles.
• High-entropy datasets facilitate better understanding and application of underlying physics models in imaging.
• Strategic selection of diverse training data is key to developing robust and interpretable DL solutions for computational imaging.