ZeroDefect YOLOv5 Nano β€” v1.0

Aircraft Skin Defect Detection (Edge-Deployable)

A YOLOv5 Nano object detection model trained to identify surface defects on aircraft skin. Optimized for edge deployment on the Luckfox Pico Max (RV1106 SoC, 0.5 TOPS NPU).


Model Formats

Format File Size Use Case
PyTorch best.pt ~3.8 MB Training, fine-tuning, desktop inference
ONNX best.onnx ~7.2 MB Cross-platform CPU/GPU inference
RKNN best.rknn ~2.6 MB RV1106 NPU hardware-accelerated inference

Classes (7)

ID Class Description
0 crack Structural cracks, fatigue lines
1 dent Impact dents, surface deformations
2 corrosion Rust, chemical wear, oxidation
3 scratch Superficial scrapes, paint scratches
4 paint-peel Flaking paint, coating degradation
5 missing-head Missing rivet heads or fasteners
6 defect Generic surface defect anomalies

Evaluation Results

Tested on 2,139 validation images (2,215 with detect.py)

Class Precision Recall mAP@0.5 mAP@0.5:0.95
all 0.435 0.355 0.323 0.238
missing-head 0.404 0.640 0.528 0.350
dent 0.576 0.448 0.447 0.372
scratch 0.509 0.397 0.363 0.356
paint-peel 0.395 0.312 0.312 0.250
crack 0.483 0.262 0.283 0.220
corrosion 0.308 0.227 0.172 0.072
defect 0.374 0.199 0.154 0.044

Inference Speed: 2.8 ms/image (RTX 3050, batch 16) | 7.8 ms/image (single image)

Confusion Matrix

Confusion Matrix

Precision-Recall Curve

PR Curve

F1-Confidence Curve

F1 Curve

Sample Detections

See the sample_detections/ folder for annotated detection results on validation images.


Training Details

  • Architecture: YOLOv5 Nano (1.77M parameters, 4.2 GFLOPs)
  • Epochs: 150
  • Image Size: 640x640
  • Batch Size: 16
  • Dataset: 22k+ train / 2.2k+ val images merged from 36 COCO-format Roboflow datasets
  • GPU: NVIDIA RTX 3050 6GB Laptop
  • Optimizer: SGD with cosine LR schedule
  • Augmentation: Standard YOLOv5 mosaic, mixup, HSV jitter

Quick Start: ONNX Inference

pip install onnxruntime opencv-python numpy
python infer_onnx.py --model best.onnx --image test.jpg --conf 0.25

Hardware Deployment (RV1106 NPU)

# Push to board
adb push best.rknn /userdata/

# Run on board (after cross-compiling rknn_model_zoo demo)
./rknn_yolov5_demo /userdata/best.rknn test.jpg

License

MIT License

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Evaluation results

  • Precision on Merged Aircraft Skin Defect & Corrosion Dataset
    self-reported
    0.435
  • Recall on Merged Aircraft Skin Defect & Corrosion Dataset
    self-reported
    0.355
  • mAP@0.5 on Merged Aircraft Skin Defect & Corrosion Dataset
    self-reported
    0.323
  • mAP@0.5:0.95 on Merged Aircraft Skin Defect & Corrosion Dataset
    self-reported
    0.238