README.md

---
license: mit
language:
- en
tags:
- object-detection
- computer-vision
- yolov8
- defect-detection
- manufacturing
- industrial-inspection
- ultralytics
pipeline_tag: object-detection
base_model:
- Ultralytics/YOLOv8
---

# 🏭 Factory Defect Guard — YOLOv8 Industrial Defect Detection

Multi-domain industrial defect detection model trained on 29,000+ images across steel surfaces, PCBs, and industrial components. Detects **17 defect classes** in a single forward pass.

| Metric | Value |
|---|---|
| mAP@0.5 | **83.0%** (V6_MC) |
| mAP@0.5:0.95 | 56.4% |
| Precision | 78.8% |
| Recall | 72.2% |
| Model size | 22.5 MB |
| Input size | 640×640 |

---

## 🔍 Defect Classes (17)

**Steel Surface (NEU Dataset)**
`crazing` · `inclusion` · `patches` · `pitted_surface` · `rolled_in_scale` · `scratches`

**PCB Defects**
`pcb_missing_hole` · `pcb_mouse_bite` · `pcb_open_circuit` · `pcb_short` · `pcb_spur` · `pcb_spurious_copper`

**Industrial Components (MVTec-derived)**
`metal_nut_defect` · `screw_defect` · `transistor_defect` · `tile_defect` · `cable_defect`

---

## 🚀 Quick Start

```python
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Load model
model_path = hf_hub_download(
    repo_id  = "negi3961/factory-defect-guard",
    filename = "best_v6_mc.pt"   # MC Dropout version — best accuracy
)
model = YOLO(model_path)

# Run inference
results = model.predict("your_image.jpg", conf=0.25)
results[0].show()

# Get detections
for box in results[0].boxes:
    cls  = int(box.cls)
    conf = float(box.conf)
    name = model.names[cls]
    print(f"{name}: {conf:.2f}")
```

---

## 📦 Model Files

| File | Description | mAP@0.5 |
|---|---|---|
| `best_v6_mc.pt` | **Recommended** — V6 fine-tuned with MC Dropout | **0.830** |
| `best.pt` | V6 base model | 0.796 |

Use `best_v6_mc.pt` for production. `best.pt` is kept for reproducibility.

---

## 🗂️ Training Details

### Datasets Used
| Dataset | Domain | Images |
|---|---|---|
| NEU Surface Defect Database | Steel surface | ~1,800 |
| PCB Defect (akhatova) | PCB original | ~1,600 |
| PCB Dataset (nakul8820) | PCB augmented | ~2,000 |
| PCB Defect (norbertelter) | PCB YOLO format | ~10,668 |
| MVTec AD subset | Industrial objects | ~428 |
| Magnetic Tile Defects | Tile surface | ~2,688 |
| Surface Defect (yidazhang07) | Mixed | ~4,194 |
| **Total** | | **~29,354** |

### Training Config (V6)
```yaml
model:     YOLOv8s
epochs:    60
imgsz:     640
batch:     16
optimizer: AdamW
lr0:       0.0001
mosaic:    1.0
mixup:     0.2
patience:  20
platform:  Kaggle GPU (Tesla T4)
```

### Training Progression
| Run | Epochs | mAP@0.5 | Notes |
|---|---|---|---|
| V5 | 43 | 0.7477 | Initial training |
| V6 | 60 | 0.7960 | Full run, AdamW |
| V6_MC | +fine-tune | **0.8300** | MC Dropout added |

---

## 📊 Per-Class mAP@0.5

| Class | mAP@0.5 |
|---|---|
| `tile_defect` | 99.5% |
| `pcb_missing_hole` | 99.3% |
| `pcb_short` | 95.5% |
| `pcb_open_circuit` | 90.7% |
| `patches` | 91.6% |
| `pcb_spurious_copper` | 91.1% |
| `pcb_mouse_bite` | 81.8% |
| `metal_nut_defect` | 85.5% |
| `inclusion` | 81.3% |
| `scratches` | 80.7% |
| `cable_defect` | 82.3% |
| `rolled_in_scale` | 57.4% |
| `screw_defect` | 56.8% |
| `transistor_defect` | 54.0% |
| `crazing` | 48.9% |

> **Note:** `crazing` is the hardest class — subtle surface texture variation makes it difficult to detect. `tile_defect` achieves near-perfect accuracy due to strong visual contrast.

---

## 🛠️ Custom Inference Pipeline

```python
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
import cv2

CLASSES = [
    'crazing', 'inclusion', 'patches', 'pitted_surface',
    'rolled_in_scale', 'scratches', 'pcb_missing_hole',
    'pcb_mouse_bite', 'pcb_open_circuit', 'pcb_short',
    'pcb_spur', 'pcb_spurious_copper', 'metal_nut_defect',
    'screw_defect', 'transistor_defect', 'tile_defect', 'cable_defect'
]

model_path = hf_hub_download("negi3961/factory-defect-guard", "best_v6_mc.pt")
model = YOLO(model_path)

def inspect(image_path, conf_threshold=0.25):
    results = model.predict(image_path, conf=conf_threshold, verbose=False)
    detections = []
    for box in results[0].boxes:
        detections.append({
            "class": CLASSES[int(box.cls)],
            "confidence": round(float(box.conf), 3),
            "bbox": box.xyxy[0].tolist()
        })
    return detections

print(inspect("surface_sample.jpg"))
```

---

## 📋 Requirements

```
ultralytics>=8.0.0
huggingface_hub
torch>=2.0.0
Pillow
opencv-python
```

---

## ⚠️ Limitations

- Model trained on specific public benchmark datasets — real factory images may need fine-tuning
- `crazing` and `transistor_defect` classes have lower accuracy (~49–54%) and may produce false negatives on ambiguous textures
- Optimized for 640×640 input; very small defects on high-resolution industrial cameras may need tiling

---

## 🔗 Links

- **GitHub:** [github.com/chandanNegi39671/factory-defect-guard](https://github.com/chandanNegi39671/factory-defect-guard) 
- **Training Notebook:** Kaggle (YOLOv8s, Tesla T4) (https://www.kaggle.com/code/negi1586/notebookce295c43f7)

---

## 👤 Author

**Negi** — ML Engineer  
HuggingFace: [@negi3961](https://huggingface.co/negi3961)