Anomaly Detection YOLOv8s — V11

Production-grade anomaly detection model for underground gas valve well inspection. Identifies 6 types of structural and environmental anomalies in inspection photos.

Model Performance

Metric	Value
mAP50	65.72%
mAP50-95	50.91%
Model	YOLOv8s
Parameters	11.1M
Input Size	640×640
Training Images	1,512 (VLM-annotated)
Fine-tuned from	V9 champion model

Per-class Performance (on V2s val set, 201 images)

Class	mAP50	mAP50-95
Water Accumulation (积水)	84.1%	65.1%
Water Seepage (渗水)	65.2%	47.6%
Corrosion / Rust (腐蚀生锈)	61.9%	45.9%
Coating Damage (涂层损坏)	39.7%	24.2%
Wall Crack (墙体裂缝)	69.3%	55.3%
Fog / Condensation (雾气结露)	74.1%	67.4%

Training History

Version	mAP50	Training Images	Key Change
V2s	48.2%	512	Baseline, VLM-annotated
V6	55.71%	1,062	VLM-only breakthrough, fine-tuned from V2s
V7	56.0%	2,779	Pseudo-labels added — no improvement
V8	60.51%	1,464	Fine-tuned from V6, more VLM data
V9	63.72%	1,463	Fine-tuned from V8, more VLM data
V11	65.72%	1,512	Fine-tuned from V9, consistent val split

Key Insights

Iterative fine-tuning works: Each generation fine-tuned from the previous best shows consistent gains (+8.6%, +5.3%, +3.1%)
Pseudo-labels don't help: V7 (2,779 images with pseudo-labels) performed identically to V6 (1,062 VLM-only)
Val split matters: V10 used a different val split and regressed; V11 with V9's split confirmed +3.1% gain
Weak classes need focused data: water_seepage gained +10.4% with targeted VLM annotations; coating_damage still lags at 39.7%

Quick Start

pip install ultralytics

from ultralytics import YOLO

model = YOLO("best_v11.pt")
results = model.predict(source="inspection_photo.jpg", conf=0.3)

for r in results:
    for box in r.boxes:
        cls = int(box.cls[0])
        conf = float(box.conf[0])
        print(f"Anomaly: {model.names[cls]}, Confidence: {conf:.2f}")

Model Files

File	Format	Size
`best_v11.pt`	PyTorch	21.5 MB
`best_v11.onnx`	ONNX	42.7 MB
`best_v11.torchscript`	TorchScript	42.8 MB
`best_v11.mlpackage/`	CoreML	42.7 MB

Anomaly Classes & Severity

Class	Severity	Description
Water Accumulation	Medium	Standing water at well bottom
Water Seepage	High	Water seeping through walls/joints
Corrosion / Rust	High	Metal rust, pipe corrosion
Coating Damage	Medium	Peeling paint, coating deterioration
Wall Crack	Critical	Structural cracks in walls
Fog / Condensation	Low	Condensation, moisture in air

Dataset Statistics (861,367 images analyzed)

94.5% of images have at least one anomaly
Corrosion/Rust: 77.3% prevalence
Coating Damage: 67.7% prevalence
Wall Crack: 59.6% prevalence
Water Seepage: 46.1% prevalence
Water Accumulation: 44.4% prevalence
Fog Condensation: 11.0% prevalence

Companion Model

For valve detection (gate, globe, ball, other), see our Valve Detection Model (mAP50=92.95%).

lg227210
/

anomaly-detection-yolov8s