Skip to content

Navigation Menu

Sign in
Sign up

Latest commit

History

8 Commits

Folders and files

NameName
Last commit message
Last commit date

Repository files navigation

πŸŒˆπŸ“Έ OBF-Design

DOI:10.1109/STSIVA66383.2025.11156301

This framework learns optimal Gaussian Optical Bandpass Filters that transform high-dimensional spectral signatures into compact, information-rich representations 🌟

Deep Gaussian Optical Bandpass Filter Design for Fermentation Index Estimation in Cocoa Beans πŸ”¬πŸ’«

A powerful deep learning framework for learnable Optical Bandpass Filter Design! 🎯 Transform spectral signatures into meaningful insights with AI-optimized Gaussian filters πŸ“Šβœ¨

🎨 Core Innovation

FilterDesign is the ⭐ core ⭐ of this work!
It learns Gaussian optical filters that reduce spectral dimensionality while preserving the most discriminative features 🌟

πŸ”¬ Technical approach:

  • πŸ“‘ Processes raw spectral data (hundreds to thousands of bands)
  • 🧠 Learns optimal Gaussian filter parameters (ΞΌ, Οƒ)
  • βœ‚οΈ Reduces dimensionality to physically feasible spectral bands
  • 🎯 Achieves state-of-the-art performance with significantly fewer spectral inputs ✨

πŸš€ Quick Start

πŸƒβ€β™€οΈ Run FilterDesign with 6 Gaussian filters

python deep_learning.py --mode filter_design --learned-bands 6 --epochs 10

πŸ”„ Compare with baseline (all spectral bands, no filtering)

python deep_learning.py --mode baseline --epochs 10

🎲 Experiment with different filter counts

python run.py # Tests 3, 6, 11 filters automatically

πŸ› οΈ Available Methods

Mode Description
filter_design 🌟 Primary method - Learnable Gaussian optical filters
band_selection Learnable binary spectral band selection
binary_band_selection Hard binary spectral band selection
baseline Standard full-spectrum approach (no filtering)

πŸ—οΈ Technical Architecture

  • πŸ§ͺ FilterDesign: Gaussian optical filters with learnable ΞΌ & Οƒ
  • πŸ€– Multiple Backbones: SpectraNet, CNN, LSTM, Transformer, SpectralFormer
  • πŸ“ˆ Two-Stage Training: Joint optimization + filter parameter freezing
  • πŸŽ›οΈ FWHM Constraints: Bandwidth limited to physically feasible range (8.25–41.25 nm)

πŸ“Š Performance Benefits

  • 🎯 Spectral Efficiency: Drastically reduces dimensionality with minimal performance loss
  • πŸ’‘ Interpretability: Highlights relevant spectral regions linked to the task
  • πŸ”§ Modularity: Compatible with diverse neural architectures
  • πŸ“ˆ Validated: Demonstrated superiority over state-of-the-art band selection methods

⭐ Citation

If you find the OBF-Design useful in your research, please consider citing:

@INPROCEEDINGS{11156301,
 author={Diaz-Delgado, Laura C. and Monroy, Brayan and Bacca, Jorge and Arguello, Henry},
 booktitle={2025 XXV Symposium of Image, Signal Processing, and Artificial Vision (STSIVA)}, 
 title={Deep Gaussian Optical Bandpass Filter Design for Fermentation Index Estimation in Cocoa Beans}, 
 year={2025},
 volume={},
 number={},
 pages={1-5},
 keywords={Optical filters;Filters;Biomedical optical imaging;Optical design;Estimation;Optical computing;Optical imaging;Fermentation;Optical sensors;Optical signal processing;fermentation;cocoa;deep learning;spectral imaging;filter design},
 doi={10.1109/STSIVA66383.2025.11156301}}

Bringing the magic of learnable optics to spectral analysis πŸ’–

🌸 Keep learning, keep filtering! 🌸

About

Deep Gaussian Optical Bandpass Filter Design for Fermentation Index Estimation in Cocoa Beans | STSIVA 2025 |

Topics

Resources

Stars

5 stars

Watchers

1 watching

Forks

Contributors

Languages

AltStyle γ«γ‚ˆγ£γ¦ε€‰ζ›γ•γ‚ŒγŸγƒšγƒΌγ‚Έ (->γ‚ͺγƒͺγ‚ΈγƒŠγƒ«) /