Digital hologram for data augmentation in learning-based pattern classification.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Learning-based pattern classification relies heavily on large, diverse datasets.
• Digital holography captures complex wavefront information but often yields limited labeled data.
• Existing data augmentation techniques may not fully exploit the rich information in holographic data.

Purpose of the Study:

• To introduce a novel data augmentation method using numerical focusing of digital holography.
• To enhance the performance of learning-based pattern classification by increasing the accessible labeled data.
• To leverage full-field wavefront information and 3D diffraction characteristics for improved classification.

Main Methods:

• Digital holographic data augmentation (DHDA) was developed based on numerical focusing.
• A complex pattern diffraction approach was employed to access full-field wavefront data.
• The method aimed to minimize confusion in the diffraction regime to preserve sample information.

Main Results:

• DHDA significantly increased the amount of accessible labeled data.
• The method provided diverse three-dimensional diffraction characteristics.
• Pattern classification accuracy improved from 90.9% without DHDA to 95.1% with DHDA.

Conclusions:

• Digital holographic data augmentation is an effective technique for enhancing learning-based pattern classification.
• DHDA complements the data manifold by providing richer 3D diffraction information.
• The proposed method offers a promising approach for improving classification performance in data-scarce holographic applications.