nitishmehta/DeepSolarEye-cleaning-quality

DeepSolarEye - Solar Panel Cleaning Quality & Vendor Rating

🔬 Overview

A complete pipeline for assessing solar rooftop panel cleaning quality and rating cleaning vendor performance. Based on the architecture from:

"DeepSolarEye: Power Loss Prediction and Weakly Supervised Soiling Localization via Fully Convolutional Networks for Solar Panels"
Mehta et al., 2017 (arxiv:1710.03811)

📦 What's Included

File	Description
model.pth	Trained model weights (49MB)
deepsolar_config.json	Model config with classes & quality ratings
inference.py	Single-image prediction script
auto_label.py	Auto-label unlabeled images using SigLIP zero-shot + vendor report
train.py	Full training script (data generation + training + Hub push)
DeepSolarEye_Training.ipynb	Google Colab notebook — train on free T4 GPU

🚀 Quick Start: 3-Step Pipeline

Step 1: Auto-Label Your Images (if unlabeled)

Uses Google's SigLIP model to automatically classify your solar panel images — no training data needed.

pip install transformers torch pillow tqdm

python auto_label.py \
    --input_dir /path/to/your/solar/images/ \
    --output_dir /path/to/labeled/ \
    --vendor_name "SolarClean Pro"

Output:

• labeled/train/ — Auto-labeled images organized by class (ready for training)
• labeled/review/ — Low-confidence images for manual review
• vendor_report.json — Cleaning quality grade (A+ to F)

Step 2: Train on Your Data (Google Colab)

1. Open the Colab notebook
2. Upload your labeled images (zip)
3. Train on free T4 GPU (50 epochs, ~30 min)
4. Push trained model to Hub

Step 3: Rate a Cleaning Vendor

python auto_label.py \
    --input_dir /photos/after_vendor_cleaning/ \
    --output_dir /reports/vendor_abc/ \
    --vendor_name "ABC Solar Cleaning"

Sample Output: ``` 🏢 VENDOR CLEANING QUALITY REPORT

Vendor: ABC Solar Cleaning Grade: B Panels Inspected: 150

📊 Quality Score: 3.67/5.0 ✅ Clean Panels: 45.3% ✅ Acceptable (Clean + Light-Dust): 72.0%

⚠️ Issues Found: - Heavy-Dust: 12 panels - Bird-Dropping: 8 panels

💰 Recommendation: Good cleaning. Minor issues to address.

## 🏗️ Architecture

**ResNet-18 + BiDIAF (Bi-Directional Input-Aware Fusion)** — 12.8M parameters

Input Image (224×224×3) │ ├── ResNet Stem (conv→bn→relu→maxpool) ├── Stage 1 (64ch) + BiDIAF Block 1 + Spatial Dropout ├── Stage 2 (128ch) + BiDIAF Block 2 + Spatial Dropout
├── Stage 3 (256ch) + BiDIAF Block 3 + Spatial Dropout ├── Stage 4 (512ch) + BiDIAF Block 4 + Spatial Dropout ├── Global Average Pooling └── FC Classifier → 6 classes → Quality Rating (1-5⭐)

**Key design decisions from the paper:**
- BiDIAF reinforces input image at every spatial level — preserves color/texture features critical for soiling
- Random weight init (not ImageNet) — learns domain-specific features
- No color augmentation — color IS the signal for soiling detection
- Inverse class-frequency weighted loss — handles imbalanced cleaning data

## 📊 Classes & Quality Ratings

| Class | Rating | Stars | Vendor Action |
|-------|--------|-------|---------------|
| Clean | 5/5 | ⭐⭐⭐⭐⭐ | Approve payment |
| Light-Dust | 4/5 | ⭐⭐⭐⭐☆ | Acceptable, note for next visit |
| Bird-Dropping | 3/5 | ⭐⭐⭐☆☆ | Deduct, request spot-clean |
| Partial-Shadow | 3/5 | ⭐⭐⭐☆☆ | Flag obstruction for removal |
| Heavy-Dust | 2/5 | ⭐⭐☆☆☆ | Reject, request re-cleaning |
| Snow-Covered | 1/5 | ⭐☆☆☆☆ | Immediate attention required |

**Vendor Grading Scale:**
| Avg Score | Grade | Action |
|-----------|-------|--------|
| ≥ 4.5 | A+ | Excellent — approve full payment |
| ≥ 4.0 | A | Great — approve payment |
| ≥ 3.5 | B | Good — minor issues to discuss |
| ≥ 3.0 | C | Acceptable — needs improvement |
| ≥ 2.0 | D | Poor — deduct payment |
| < 2.0 | F | Unacceptable — request re-clean |

## 💻 Usage: Single Image Inference

```python
python inference.py --image panel_photo.jpg

Or in Python:

from inference import load_model, predict

model = load_model()  # Downloads from Hub automatically
result = predict(model, "panel.jpg")

print(f"Condition: {result['predicted_class']}")
print(f"Rating: {result['quality_rating']}/5")
print(f"Confidence: {result['confidence']:.1%}")

📈 Training on Your Own Data

Option A: Google Colab (recommended)

• Open DeepSolarEye_Training.ipynb in Colab
• Free T4 GPU, ~30 min for 50 epochs

Option B: HF Jobs (requires Pro)

# After getting HF Pro subscription
pip install huggingface_hub
python train.py

Option C: Local GPU

pip install torch torchvision transformers datasets evaluate accelerate scikit-learn
python train.py  # Modify DATA_DIR to point to your data

🔗 References

• DeepSolarEye (WACV 2018) — Core architecture
• CNN Fault Detection for PV Panels — Multi-class PV classification
• PV-DDPM — Synthetic defect generation for PV

📝 License

MIT