High-generalization deep sparse pattern reconstruction: feature extraction of speckles using self-attention armed...

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Image Processing

Background:

• Light scattering poses a significant challenge in various scientific and imaging applications.
• Deep learning has recently shown promise in addressing speckle reconstruction problems.

Purpose of the Study:

• To investigate the feature extraction and generalization capabilities of deep learning networks for sparse pattern reconstruction.
• To develop an advanced neural network for enhanced speckle reconstruction.

Main Methods:

• Development of the "one-to-all" self-attention armed convolutional neural network (SACNN).
• Utilizing SACNN to extract local and global speckle properties from diverse sparse patterns, including unseen diffusers and untrained positions.
• Quantitative performance analysis using scientific indicators like Pearson correlation coefficient, structural similarity measure, and Jaccard index.

Main Results:

• SACNN demonstrated superior performance compared to traditional convolutional neural networks, with over 10% improvement in key metrics.
• The network successfully reconstructed features significantly beyond the memory effect range (75x for a 120 grits diffuser).
• SACNN exhibits strong generalization abilities across different sparse patterns and detection scenarios.

Conclusions:

• The developed SACNN effectively addresses light scattering challenges in sparse pattern reconstruction.
• This method enhances the field of view and depth of field for imaging through complex scatters.
• The findings have significant implications for applications such as deep tissue imaging.