Multi-Class Classification and Multi-Output Regression of Three-Dimensional Objects Using Artificial Intelligence...

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Digital holography captures 3D information, crucial for AI-driven visual perception.
• Processing holographic data for AI applications requires robust machine learning approaches.
• Supervised learning offers a framework for extracting meaningful insights from complex holographic datasets.

Purpose of the Study:

• To demonstrate digital holographically sensed 3D data processing for AI-based vision.
• To evaluate multiple machine learning methods for 3D object classification and regression using holographic data.
• To compare the performance of different data representations (holograms, intensity-phase images, phase-only images) in supervised learning tasks.

Main Methods:

• Utilized three datasets: sensed holograms, concatenated intensity-phase images, and phase-only images.
• Employed supervised learning for multi-class classification and multi-output regression tasks.
• Compared performance of machine learning classifiers (AUC values) and regressors (EV regression score) including CNN, MLP, and RF.

Main Results:

• Machine learning classifiers outperformed CNN in AUC for holograms and whole information datasets.
• CNN showed superior performance on phase-only image datasets.
• MLP regressor achieved a stable prediction (EV score 0.00) across datasets, while RF excelled on whole information data (EV score 0.01).

Conclusions:

• Different machine learning models and holographic data representations yield varying performance in 3D object analysis.
• The choice of model and data type is critical for optimizing AI-based vision tasks using digital holography.
• This work validates the efficacy of digital holographic data processing for advanced AI applications.