Dictionary Learning Phase Retrieval from Noisy Diffraction Patterns.

Area of Science:

• Computational Imaging
• Signal Processing
• Machine Learning

Background:

• Image phase retrieval is crucial for reconstructing complex-valued images from amplitude measurements.
• Existing methods struggle with heavily noisy observations, such as Poissonian or Gaussian noise.
• Sparse coding offers a promising approach for image reconstruction in challenging conditions.

Purpose of the Study:

• To develop a novel algorithm for robust image phase retrieval under significant noise.
• To integrate dictionary learning with sparse regression for enhanced image reconstruction.
• To introduce the Dictionary Learning Phase Retrieval (DLPR) algorithm.

Main Methods:

• The algorithm employs the alternating projection framework.
• Image estimation is reformulated as sparse regression in the complex domain.
• A complex domain dictionary is learned from data using matrix factorization with sparsity constraints.

Main Results:

• The proposed DLPR algorithm jointly learns the dictionary and reconstructs the target image.
• Experiments demonstrate DLPR's effectiveness on simulated and real complex image data.
• DLPR shows noticeable advantages over state-of-the-art competitors in noisy scenarios.

Conclusions:

• DLPR provides a powerful new tool for image phase retrieval, especially in noisy environments.
• The joint dictionary learning and reconstruction approach enhances robustness and performance.
• This method offers significant improvements for applications requiring accurate complex image recovery.