prath0029/Zerodefect-1.0-bucket

ZeroDefect — YOLOv5 Nano (v2.0 Class-Balanced)

Aircraft Skin & Automotive Component Defect Detection (Edge-Deployable)

A YOLOv5 Nano object detection model trained on a class-balanced merged dataset of aircraft skin and automotive component defects. This is version 2.0 of the ZeroDefect model, trained with:

• AdamW optimizer + Cosine LR decay for stable convergence
• Class-balanced sampling across 7 defect categories
• Multi-scale training (random image size between 0.5x–1.5x)
• Label smoothing (0.05) for better generalization
• 150 epochs fine-tuning from a high-quality pretrained base

Detected Defect Categories (7 Classes)

ID	Class	Description
0	crack	Structural cracks, fatigue lines, stress fractures
1	dent	Impact dents, surface deformations
2	corrosion	Rust, chemical wear, oxidation spots
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 anomalies

---

Training & Validation Metrics

Run	Epochs	mAP@0.5	mAP@0.5:0.95	Precision	Recall
Balanced v2.0 (this repo, best epoch=101)	150	0.30994	0.22613	0.37777	0.34799
Nano v1.0 (baseline, epoch 149)	150	0.31962	0.23492	0.42944	0.35807

Note: The class-balanced v2.0 run achieves comparable mAP scores to the nano v1.0 baseline while training on a harder, class-balanced dataset. The lower absolute mAP reflects that the dataset had its class distribution enforced (harder to overfit to the dominant "defect" class), resulting in more balanced per-class detection capability.

Training Configuration

• Model Architecture: YOLOv5 Nano backbone (~1.76M parameters, 4.2 GFLOPs)
• Dataset: Merged aircraft skin defect + corrosion dataset (~30k images, 7 classes)
• Optimizer: AdamW | Cosine LR decay
• Label Smoothing: 0.05
• Early Stopping: patience=30 (best epoch: 101/150)
• Hardware: NVIDIA RTX 3050 Laptop GPU (6 GB VRAM)

---

Repository Contents

File	Description	Size
best.pt	PyTorch model weights (YOLOv5 format)	~3.8 MB
best_balanced.onnx	ONNX export (opset 18, simplified)	~7.5 MB
best.rknn	RKNN binary for RV1106 NPU (INT8, prior run)	~2.6 MB
data.yaml	Dataset configuration	<1 KB
infer_onnx.py	Python ONNX inference helper	~8 KB
evaluation/	Training curves, confusion matrix, PR/F1 plots	—

---

Quick Start: ONNX Inference (Python)

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

Hardware Deployment (Luckfox Pico Max / RV1106 NPU)

The best.rknn model is compiled for the RV1106 NPU (INT8 quantized):

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

# Build C++ inference runner from rknn_model_zoo
./build-linux.sh -t rv1106 -a armv7l -d yolov5

# Run inference on board
./rknn_yolov5_demo /userdata/best.rknn test.jpg

---

License

Released under the MIT License. See LICENSE for details.