| # EL Defect Detection — Training Package |
|
|
| Train a **U-Net++ with EfficientNet-B4 encoder** for solar cell EL defect segmentation. |
|
|
| ## Quick Start (RTX 4060 / any CUDA GPU) |
|
|
| ```bash |
| # 1. Clone this repo |
| git clone https://huggingface.co/nithishbasireddy/el-defect-training |
| cd el-defect-training |
| |
| # 2. Install dependencies |
| pip install -r requirements.txt |
| |
| # 3. Verify setup |
| python test_setup.py |
| |
| # 4. Train (auto-downloads E-SCDD dataset) |
| python train.py |
| ``` |
|
|
| Training takes ~2-3 hours on RTX 4060. Model saves to `output/best_model.pth`. |
|
|
| ## After Training |
|
|
| ```bash |
| # Run the Streamlit app with your trained model |
| streamlit run app.py |
| # → Set model path to "output/best_model.pth" in sidebar |
| ``` |
|
|
| ## Architecture |
|
|
| - **Model**: U-Net++ + EfficientNet-B4 + scSE attention (20.9M params) |
| - **Dataset**: E-SCDD (903 images, 512×512, 30→5 class remap) |
| - **Loss**: 0.5 × Dice + 0.5 × Focal (γ=2.0) — handles severe class imbalance |
| - **Optimizer**: AdamW with differential LR (encoder 1e-4, decoder 5e-4) |
| - **AMP**: Mixed precision for 8GB VRAM GPUs |
|
|
| ## Classes (5) |
|
|
| | ID | Name | E-SCDD Labels | Color | |
| |----|------|---------------|-------| |
| | 0 | background | 0-8, 21-24, 29 | Black | |
| | 1 | busbar | 9 | Green | |
| | 2 | crack | 10, 14 | Blue | |
| | 3 | dark/inactive | 11, 17, 20 | Red | |
| | 4 | other_defect | 12,13,15,16,18,19,25-28 | Yellow | |
| |
| ## Why U-Net++ over U-Net? |
| |
| U-Net++ has **dense skip connections** between encoder and decoder. For thin structures like cracks (often <10px wide), these nested connections preserve fine details that plain U-Net's single skip connections lose. Validated in medical imaging (thin vessels) and industrial inspection. |
| |
