| license: other | |
| tags: | |
| - object-detection | |
| - pcb-defect-detection | |
| - yolov8 | |
| - k210 | |
| - edge-ai | |
| - fine-tuned | |
| # Project_final — PCB Assembly Defect Detection (v2) | |
| Trained for **PS82 — Low-cost Optical Inspection for PCB Assembly**. | |
| Base model: **YOLOv8n**, fine-tuned in two phases: | |
| 1. **Phase 1 (30 ep):** Backbone frozen, head-only training with lr=0.01 | |
| 2. **Phase 2 (170 ep):** Full fine-tuning with lr=0.001, cosine decay, aggressive augmentation | |
| Input resolution: **320×320** (K210-safe) | |
| Target deployment: **Sipeed Maix Bit (Kendryte K210)**, via ONNX → NNCase INT8 `.kmodel`. | |
| ## Classes | |
| ['solder_bridge', 'missing_component', 'component_misalignment', 'cold_solder_joint'] | |
| ## Training Improvements (v2) | |
| - Transfer learning from saved checkpoint | |
| - Two-phase training (freeze backbone → full fine-tune) | |
| - Offline augmentation for underrepresented classes | |
| - Cosine LR schedule with warmup | |
| - Label smoothing (0.05) | |
| - Aggressive data augmentation (mosaic, mixup, copy-paste, perspective) | |
| - AdamW optimizer | |
| See `dataset/ATTRIBUTION.md` for source dataset licensing. | |