walterfan1322/2026raman

Sudan Red detection via Raman spectroscopy using SE-ResNet deep learning # 2026raman — Sudan Red Detection via Raman Spectroscopy Deep learning system for detecting Sudan Red dyes (banned food colorants) in spectroscopic samples using Squeeze-and-Excitation Residual Networks (SE-ResNet). ## Overview Classifies Raman spectra into 4 independent binary tasks — Sudan Red 1, 2, 3, 4 — plus a hierarchical presence detector. The project documents a full 13-version evolution from baseline to production-ready architecture. ## Best Results (v11_arch — SEResNetLite) | Class | Accuracy | Recall | F1 | |-------|----------|--------|-----| | Sudan Red 1 | 83.3% | 78.6% | 73.3% | | Sudan Red 2 | 93.8% | 80.0% | 84.2% | | Sudan Red 3 | 95.8% | 81.8% | 90.0% | | Sudan Red 4 | 97.9% | 92.3% | 96.0% | | **All-or-None** | **81.3%** | — | — | | **Exclusive Top-1** | **87.5%** | — | — | ## Architecture **SEResNetLite** (v11_arch): - Conv1d(1→32) → 4 residual blocks with SE attention (32→64→128→256) - Dual pooling: adaptive avg-pool + max-pool → 512-dim - Dropout (0.2) → FC(512→1) per class - He initialization, BCEWithLogitsLoss, AdamW (lr=1e-3, wd=1e-5) **Hierarchical Pipeline** (v13): - Stage 1: Sudan presence detector (positive vs. negative) - Stage 2: 4 independent classifiers on Sudan-positive samples only - Independent data splitting to prevent label bias ## Preprocessing Pipeline 1. **Baseline correction** — Zhang fit (2 passes) 2. **Smoothing** — Savitzky-Golay (window=15, polyorder=3) 3. **Interpolation** — 400–2000 cm⁻¹ → 4096 points 4. **Normalization** — min-max / SNV / area normalization ## Project Structure ``` 2026raman/ ├── SE_ResNet_4model.py # Original baseline (v1) ├── SE_ResNet_4model_v2.py ~ v13.py # Version evolution ├── SE_ResNet_4model_v11_arch.py # Best architecture ├── SE_ResNet_4model_v13.py # Latest: hierarchical pipeline ├── requirements.txt ├── PROJECT_STRUCTURE.md ├── model_versions/ # Trained model manifests │ ├── v8/ensemble_manifest.json │ ├── v9/hybrid_manifest.json │ └── v10/tta_m …