Opportunity
The patent addresses a critical challenge in object identification and classification, particularly for biological samples, where traditional methods struggle to extract meaningful information from wavefront data. Current techniques face limitations because wavefronts (e.g., from digital holograms) often lack visually interpretable features, making it difficult to classify or identify objects accurately. Additionally, biological samples are frequently incomplete, damaged, or flawed, further complicating traditional image-based recognition. Existing solutions rely heavily on manual analysis or simplistic algorithms, which are inefficient, error-prone, and cannot handle high-noise or partial data. This patent solves these problems by employing deep learning to analyze holographic wavefronts, enabling precise identification and classification even for imperfect samples.
Technology
The patent introduces a computer-implemented method and system that processes digital hologram data using a neural-network-based ensemble model. The hologram data contains phase and magnitude information, which are processed separately by two neural networks—one for magnitude features and another for phase features. These features are then combined in a concatenate unit to classify or identify the object. The innovation lies in:
1. Wavefront Analysis: The system extracts complex wavefront data (magnitude + phase) from holograms, which retains more structural information than traditional imaging.
2. Deep Learning Ensemble: The dual-network architecture leverages both magnitude and phase features, improving robustness against noise and incomplete data.
3. Interferometer Integration: The holograms can be captured using an off-axis interferometer (e.g., Mach-Zehnder setup), enabling high-quality wavefront reconstruction.
4. Fourier Transform Processing: The raw holograms are transformed into frequency domains, where noise and artifacts can be filtered before reconstruction.
This approach significantly outperforms conventional methods, achieving 99.6% accuracy in classifying biological tissues despite imperfections.
Advantages
- High Accuracy: Achieves 99.6% classification accuracy even with flawed or noisy samples.
- Robust to Imperfections: Works well with incomplete, damaged, or occluded objects.
- Versatile Applications: Applicable to biological samples, industrial defect detection, and more.
- Real-Time Processing: Optimized for integration with high-speed interferometers and computational pipelines.
- Automation: Reduces reliance on manual analysis, improving efficiency.
Applications
- Biological Sample Classification: Identifying tissues, cells, or chromosomes in medical diagnostics.
- Industrial Inspection: Detecting defects in transparent/translucent materials (e.g., glass, LCD panels).
- Environmental Monitoring: Classifying pollutants (e.g., microplastics) in water or air.
- Security & Forensics: Authenticating materials based on structural wavefront signatures.
- Holographic Microscopy: Enhancing 3D imaging for research and clinical applications.
