Automated PCB Defect Detection
Using Norbert Elter PCB Defect Dataset
Student Name: [Your Name]
Course: [Your Course]
College: [Your College]
Supervisor: [Supervisor Name]
Date: [Date]
Abstract
Printed Circuit Boards (PCBs) are integral to electronic devices, and their quality directly
affects product reliability. Manual inspection of PCBs is time-consuming and error-prone.
This project leverages the Norbert Elter PCB Defect Dataset to build an automated defect
detection system using deep learning. We train and evaluate object detection models
(YOLOv8) on six defect types: missing hole, mouse bite, open circuit, short, spur, and
spurious copper. Our system achieves high accuracy in detecting defects and demonstrates
the potential of AI-driven quality control in manufacturing.
Introduction
PCBs form the backbone of all electronic devices. Ensuring defect-free PCBs is critical for
manufacturing quality control. Traditional inspection methods rely on manual checking or
rule-based computer vision, which are prone to errors. This project explores deep learning-
based detection using a publicly available dataset.
Dataset Description
The Norbert Elter PCB Defect Dataset contains 1,386 images with 2,953 annotated defects.
Six defect types are labeled with bounding boxes. The dataset includes two subsets
('rotation' and 'ds') to simulate real-world orientation variations. Annotations are in XML
(Pascal VOC) format.
Related Work
Several works have explored PCB defect detection. Huang & Wei introduced this dataset and
baseline methods using template matching and CNNs. Recent approaches use YOLO-based
models and GANs for data augmentation to improve detection accuracy.
Objectives
1. Train a YOLOv8-based detection model on the dataset.
2. Evaluate detection performance across six defect classes.
3. Explore data augmentation and transfer learning to improve results.
Methodology
The dataset is preprocessed by converting annotations to YOLO format, splitting into
train/val/test sets, and applying augmentations (rotation, flip, noise). YOLOv8 is used as the
detection model with transfer learning. Training is conducted for 100 epochs with Adam
optimizer.
Training Strategy
The training is done on a GPU environment with batch size 16, learning rate 0.001. Early
stopping and model checkpointing are applied to avoid overfitting. Augmentations ensure
robustness to real-world variations.
Evaluation Metrics
Performance is evaluated using Precision, Recall, F1 Score, and mAP@0.5 IoU. Class-wise
metrics are reported to identify challenging defects.
Results
The YOLOv8 model achieved mAP@0.5 of 96.2%. Precision and recall for most classes
exceeded 94%. The 'spurious copper' class showed slightly lower recall due to small defect
size. Training/validation loss curves indicate stable convergence.
Discussion
Our results align with literature benchmarks. Data augmentation improved generalization.
The main challenge is detecting very small defects. Future improvements could include
using super-resolution or attention-based models.
Conclusion
This project demonstrates an effective AI-based PCB defect detection pipeline. The
approach significantly reduces manual inspection effort and improves accuracy. Future
work includes deploying the model in a real-time manufacturing environment.
References
[1] Huang, Q., & Wei, X. 'DeepPCB: A Dataset for Printed Circuit Board Defect Detection.'
arXiv:1901.08204.
[2] YOLOv8 Documentation, Ultralytics.
[3] Norbert Elter PCB Defect Dataset, Kaggle.