Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,119 @@
+---
+license: mit
+language:
+- en
+tags:
+- yolo
+- anomaly-detection
+- industrial-inspection
+- valve-well
+- computer-vision
+- object-detection
+library_name: ultralytics
+pipeline_tag: object-detection
+---
+
+# Anomaly Detection YOLOv8s — Underground Gas Valve Well Inspection
+
+## Overview
+
+A YOLOv8s-based object detection model for identifying and localizing anomalies in underground gas valve well inspection photos. This model is trained on VLM-annotated real-world inspection images and detects 6 types of anomalies commonly found in valve wells.
+
+## Model Performance
+
+| Metric | Value |
+|--------|-------|
+| **mAP50** | 0.312 |
+| **mAP50-95** | 0.194 |
+| **Precision** | 0.624 |
+| **Recall** | 0.306 |
+| **Model Size** | 22.5 MB (PyTorch) / 44.7 MB (ONNX) |
+| **Parameters** | 11.1M |
+| **GFLOPs** | 28.4 |
+
+## Anomaly Classes
+
+| ID | Class | mAP50 | Recall | Description |
+|----|-------|-------|--------|-------------|
+| 0 | water_accumulation | 0.577 | 0.704 | Standing water on the floor |
+| 1 | water_seepage | 0.140 | 0.115 | Water traces/seepage on walls |
+| 2 | corrosion_rust | 0.223 | 0.059 | Rust on metal parts |
+| 3 | coating_damage | 0.007 | 0.000 | Peeling or damaged coating |
+| 4 | wall_crack | 0.650 | 0.600 | Cracks in walls or structure |
+| 5 | fog_condensation | 0.202 | 0.200 | Fog or condensation on lens |
+
+## Training Details
+
+- **Base Model**: YOLOv8s (pretrained on COCO)
+- **Dataset**: 213 VLM-annotated images (173 train / 40 val)
+- **Annotations**: 505 bounding boxes across 6 anomaly classes
+- **Image Size**: 640×640
+- **Batch Size**: 8
+- **Optimizer**: AdamW (lr=0.001, momentum=0.9)
+- **Epochs**: 64 (early stopped, patience=30)
+- **Best Epoch**: 33
+- **Device**: Apple M4 (MPS)
+- **Augmentation**: Mosaic, MixUp, HSV, rotation, flip, copy-paste, random erasing
+
+## Dataset
+
+The training data consists of real underground gas valve well inspection photos collected from 701 inspection sites across Guangdong, Shaanxi, and Xinjiang provinces in China. Images are 2592×1944 resolution JPG files.
+
+Annotations were generated using a Vision-Language Model (VLM) with expert review, providing precise bounding boxes for each anomaly type.
+
+## Usage
+
+### Python (PyTorch)
+
+```python
+from ultralytics import YOLO
+
+# Load model
+model = YOLO("best.pt")
+
+# Run inference
+results = model("your_image.jpg")
+
+# Show results
+for r in results:
+    r.show()
+    print(r.boxes)
+```
+
+### ONNX
+
+```python
+import onnxruntime as ort
+import numpy as np
+from PIL import Image
+
+session = ort.InferenceSession("best.onnx")
+image = Image.open("your_image.jpg").resize((640, 640))
+input_tensor = np.array(image).transpose(2, 0, 1).astype(np.float32) / 255.0
+input_tensor = input_tensor[np.newaxis]
+
+outputs = session.run(None, {"images": input_tensor})
+```
+
+## Complete Pipeline
+
+This model is part of a complete valve well inspection pipeline:
+
+1. **Valve Detection** ([lg227210/valve-detection-yolov8s](https://huggingface.co/lg227210/valve-detection-yolov8s)) — mAP50=92.95%, 4 valve types
+2. **Anomaly Classification** — EfficientNet-B0, 74.0% accuracy, 6 anomaly types + severity
+3. **Anomaly Detection** (this model) — YOLOv8s, localizes anomalies with bounding boxes
+
+## License
+
+MIT License — Free for research and commercial use.
+
+## Citation
+
+```bibtex
+@misc{anomaly_detection_yolov8s,
+  title={Anomaly Detection YOLOv8s for Underground Gas Valve Well Inspection},
+  author={ValveAI},
+  year={2026},
+  url={https://huggingface.co/lg227210/anomaly-detection-yolov8s}
+}
+```