# YOLO-PCB

**Repository Path**: thinkinapp/YOLO-PCB

## Basic Information

- **Project Name**: YOLO-PCB
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-03-25
- **Last Updated**: 2026-03-25

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# YOLO-PCB 🔬

**A Lightweight Multi-Scale Neural Network for Accurate and Real-Time PCB Defect Detection**

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
[![Python 3.8+](https://img.shields.io/badge/Python-3.8%2B-blue.svg)](https://www.python.org/)
[![PyTorch](https://img.shields.io/badge/PyTorch-1.13%2B-red.svg)](https://pytorch.org/)

> **Paper**: Tang Jun, "YOLO-PCB: A Lightweight Multi-Scale Neural Network for Accurate and Real-Time PCB Defect Detection", IEEE Transactions on Industrial Informatics (under review)  
> **Author**: Tang Jun · Guangzhou Xinhua University · `tiger.tang@foxmail.com`

---

## Highlights

| Model | mAP@0.5 | mAP@0.5:0.95 | FPS (RTX 3090) | Params |
|-------|---------|--------------|----------------|--------|
| YOLOv8s | 89.4 | 69.7 | 108 | 11.2M |
| CDI-YOLO | 91.8 | 72.3 | 108 | 5.5M |
| **YOLO-PCB (ours)** | **94.3** | **74.8** | **156** | **4.2M** |

- **CSPPCoT backbone**: CoT self-attention inside CSP residual blocks — long-range context at < 5 M params
- **MSFE neck**: PConv spatial detail + CSPSPPF multi-receptive-field context, dual-path FPN/PAN fusion
- **PCB-SIoU loss**: Adaptive angular-cost with exponential decay for small objects — +4.4 pp AP on micro-defects (< 32×32 px)
- **Edge-ready**: 97 FPS (FP16) on NVIDIA Jetson AGX Orin

---

## Installation

```bash
git clone https://github.com/tangjun-yolo/YOLO-PCB.git
cd YOLO-PCB
pip install -r requirements.txt
```

---

## Quick Start

### Run Detection on an Image

```bash
python detect.py \
    --source path/to/pcb.jpg \
    --weights weights/best.pt \
    --conf 0.25 \
    --save-dir runs/detect
```

### Run on a Directory

```bash
python detect.py --source data/images/test/ --weights weights/best.pt
```

### Run Tests

```bash
python tests/test_yolo_pcb.py
# or with pytest
pip install pytest
pytest tests/ -v
```

---

## Training

### 1. Prepare Dataset

```bash
# DeepPCB (1386 images, 6 defect classes)
python scripts/prepare_dataset.py
```

Expected layout:
```
data/DeepPCB/
  images/
    train/   *.jpg
    val/     *.jpg
  labels/
    train/   *.txt   # YOLO format: cls cx cy w h
    val/     *.txt
```

### 2. Train

```bash
python train.py \
    --data data/configs/deeppcb.yaml \
    --epochs 300 \
    --batch 16 \
    --img-size 640 \
    --amp
```

Checkpoints are saved to `weights/`. Best model: `weights/best.pt`.

---

## Architecture

```
Input (640×640×3)
    │
    ▼
┌─────────────────────────────────────┐
│  CSPPCoT Backbone                   │
│  Stem → Stage1 (P3,80×80,128ch)    │
│       → Stage2 (P4,40×40,256ch)    │
│       → Stage3+SPPF (P5,20×20,512ch)│
└──────────────┬──────────────────────┘
               │ P3, P4, P5
               ▼
┌─────────────────────────────────────┐
│  MSFE Neck (FPN + PAN)              │
│  MSFEBlock = PConv ⊕ CSPSPPF       │
└──────────────┬──────────────────────┘
               │ N3, N4, N5
               ▼
┌────────────────────────────────────────────────────┐
│  Decoupled Heads (stride 8 / 16 / 32)              │
│  cls_branch (2×Conv) | reg_branch (2×Conv+obj)     │
└────────────────────────────────────────────────────┘
               │
               ▼
        PCB-SIoU Loss
```

---

## Defect Classes (DeepPCB)

| ID | Class | Description |
|----|-------|-------------|
| 0 | missing_hole | Absence of through-hole |
| 1 | mouse_bite | Edge notch on copper trace |
| 2 | open_circuit | Broken copper trace |
| 3 | short | Unintended copper bridge |
| 4 | spur | Extra copper protrusion |
| 5 | spurious_copper | Isolated copper residue |

---

## Repository Structure

```
YOLO-PCB/
├── models/
│   ├── modules.py      # CSPPCoT, CoTAttention, PConv, CSPSPPF, PCB-SIoU
│   └── yolo_pcb.py     # Full YOLOPCB model, backbone, neck
├── utils/
│   ├── dataset.py      # PCBDataset, augmentation, collate_fn
│   └── postprocess.py  # NMS, decode_predictions, compute_map
├── tests/
│   └── test_yolo_pcb.py # 17 unit & integration tests
├── data/
│   └── configs/
│       └── deeppcb.yaml
├── scripts/
│   └── prepare_dataset.py
├── inference/          # (reserved for TensorRT / ONNX export)
├── train.py            # Training script with AMP support
├── detect.py           # Inference CLI (image / video / webcam)
├── requirements.txt
└── README.md
```

---

## Citation

```bibtex
@article{tang2026yolopcb,
  title   = {{YOLO-PCB}: A Lightweight Multi-Scale Neural Network
             for Accurate and Real-Time {PCB} Defect Detection},
  author  = {Tang, Jun},
  journal = {IEEE Transactions on Industrial Informatics},
  year    = {2026},
  note    = {Under review}
}
```

---

## License

MIT © Tang Jun, Guangzhou Xinhua University