ESPI Mode Classification (RF & CNN)

Κύριο repo για την τελική ταξινόμηση modes ESPI. Περιλαμβάνει:

1. Random Forest (Baseline & Hybrid): Το τελικό νικητήριο μοντέλο.
2. CNN (ResNet-18): Deep Learning baseline & MC-LOBO stress-tests.
3. v6.1: Future work model (frozen backbone).

---

Locked / Verified Results (Main Thesis KPIs)

Dataset

• Final labeled samples: 3,443
• Extreme imbalance ratio: 30.9:1

Performance (Standard split)

Model	Accuracy	Macro-F1
Hybrid RF (Winner)	97.85%	95.15%
Pattern-only RF	90.15%	69.91%
CNN (reference)	93.76%	88.11%

Robustness

• LOBO Accuracy: 91.83% ± 8.9% (Robustness Protocol)
• LODO Accuracy: 66.31% ± 44.11% (Robustness Protocol)
• CNN MC-LOBO pct20: 67.68% ± 2.2% (Stress-Test)

Note: MC-LOBO is an exploratory stress-test used to evaluate domain shift difficulty, not an official thesis KPI.

---

📁 Data Availability

⚠️ The raw ESPI measurement data is not included in this repository due to size constraints.

To reproduce results:

1. Contact the author for access to the ESPI dataset (PhaseOut folders: W01/W02/W03)
2. Generate the dataset CSV:

   python src/make_espi_labels_csv.py generate --roots [your_data_paths]

3. Run training scripts as described in REPRODUCE.md

Related Repository

For ESPI image preprocessing and denoising (DnCNN-ECA), see:

• ESPI-DnCNN-ECA (if exists)

---

Quickstart

# Windows (PowerShell)
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

:: Windows (CMD)
python -m venv .venv
.\.venv\Scripts\activate.bat
pip install -r requirements.txt

# Linux/macOS
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

See REPRODUCE.md for exact commands.

Citation

If you use this code in your research, please cite:

@mastersthesis{spyridakis2026espi,
  author       = {Spyridakis, Georgios},
  title        = {Automatic Modal Classification of {ESPI} Images for Musical Instrument Quality Control},
  school       = {Hellenic Mediterranean University, Department of Music Technology and Acoustics},
  year         = {2026},
  type         = {MSc Thesis},
  address      = {Rethymno, Greece},
  note         = {Hybrid Random Forest approach achieving 97.85\% accuracy on bouzouki soundboard modal classification}
}

License

MIT License - see LICENSE for details.