🔬 PCB Defect Detection using Deep Learning

Automated Quality Control System for Printed Circuit Board Manufacturing

🎯 Project Overview

An end-to-end deep learning system for automated PCB defect detection that combines computer vision with domain expertise. This project demonstrates the practical application of AI in industrial quality control, achieving 91.2% F1-score on multi-label defect classification.

Built as a portfolio project showcasing:

🧠 Deep Learning & Computer Vision expertise
⚙️ Industrial ML system design
🔧 Electrical Engineering domain knowledge
📊 Data-driven problem solving

⚡ Key Highlights

✅ 91.2% F1-Score on DeepPCB benchmark dataset
🎯 6 Defect Types: Open circuits, shorts, mousebites, spurs, spurious copper, pin-holes
⚙️ Real-time inference: 20ms per image (50 images/second)
📈 Stable training: Dropout + Weight Decay regularization
🔄 Production-ready: Automated QA report generation
💰 ROI: 6-12 months payback period

🏗️ System Architecture

Input PCB Image (640×640px)
         ↓
    Preprocessing
         ↓
  ResNet-18 CNN Backbone
   (Transfer Learning)
         ↓
   Dropout Layer (0.5)
         ↓
  Multi-Label Classification
    (6 defect classes)
         ↓
   Sigmoid Activation
         ↓
  Confidence Scores (0-1)
         ↓
  Threshold Decision (0.5)
         ↓
Quality Control Report

📊 Performance Metrics

Overall Performance

Metric	Score	Industry Target
F1 Score	91.2%	85-95% ✅
Precision	86.2%	>80% ✅
Recall	98.3%	>95% ✅
Inference Time	20ms/image	<100ms ✅

Per-Class Performance

Defect Type	F1 Score	Notes
Open Circuit	97.7%	Excellent detection ⭐
Short Circuit	87.7%	Good performance
Mousebite	90.5%	Strong recall
Spur	85.3%	Challenging class
Spurious Copper	95.7%	Very good
Pin-hole	93.2%	Excellent precision

🎨 Visualizations

Detection Pipeline

Complete detection pipeline showing template comparison to AI classification

Results Summary

Quick summary of detection results across multiple samples

Model Predictions

Detailed model predictions with confidence scores

Performance Analysis

Training stability comparison between models

Per-class confusion matrices showing detection accuracy

Detailed precision-recall analysis by defect type

🧪 What Makes This Project Unique

1. Regularization Strategy 🎓

Implemented Dropout (0.5) + L2 Weight Decay (1e-4) to prevent overfitting:

Result: Training stability improved by 81%

2. Template-Based Quality Reports 📋

Unlike generic AI models (BLIP), we use domain-specific templates:

PCB QUALITY INSPECTION REPORT
═══════════════════════════════════════
Status: ⚠️ FAIL - HIGH Severity

DETECTED DEFECTS: 2
  • SHORT CIRCUIT (Confidence: 94%)
    → Unintended electrical connection
    → Risk: Component damage, fire hazard
  
  • OPEN CIRCUIT (Confidence: 87%)
    → Discontinuity in electrical path
    → Risk: Non-functional board

RECOMMENDATIONS:
  1. URGENT: Do NOT proceed to assembly
  2. Review etching process parameters
  3. Inspect batch for similar defects

Why This Matters: 100% technical accuracy vs. <10% with generic BLIP model

3. Precision-Recall Optimization ⚖️

Deliberately prioritized high recall (98.3%) over precision (86.2%) because:

Error Type	Business Impact
False Negative (missed defect)	Board ships to customer → Field failure → $1,000+ cost
False Positive (false alarm)	Extra 2-min inspection → $2 cost

Decision: Better to have false alarms than miss critical defects!

🛠️ Technical Implementation

Tech Stack

Core Framework:

Python 3.9+
PyTorch 2.0+
torchvision (ResNet-18)

Data Processing:

OpenCV - Image preprocessing
NumPy - Numerical operations
Pandas - Data manipulation

Visualization:

Matplotlib & Seaborn
Confusion matrices
Training curves

Dataset:

DeepPCB (1,500 PCB image pairs)
6 defect classes
640×640px resolution

Model Architecture

ResNet-18 with regularized classifier head:

Pretrained on ImageNet for transfer learning
Dropout layer (0.5) to prevent overfitting
Multi-label output for simultaneous defect detection
Sigmoid activation for independent class probabilities

Training Configuration

Epochs: 20 (with early stopping)
Batch size: 16
Learning rate: 0.001
Optimizer: Adam with weight decay (1e-4)
Loss: Binary Cross-Entropy
Scheduler: ReduceLROnPlateau

Data Augmentation:

Random horizontal/vertical flips
Random rotation (±15°)
Color jitter (brightness, contrast)

📁 Project Structure

pcb-defect-detection/
├── notebooks/
│   └── 01_EDA_exploration.ipynb
├── src/
│   └── dataset_utils.py
├── models/
│   ├── best_model.pth
│   ├── best_model_regularized.pth
│   └── .....
├── results/
│   ├── complete_training_comparison.png
│   ├── confusion_matrices_regularized.png
│   ├── per_class_performance.png
│   ├── sample_predictions.png
│   ├── IEEE_Project_Report.md
│   └── .....
├── DeepPCB/
├── requirements.txt
├── .gitignore
└── README.md

🚀 Getting Started

Prerequisites

Python 3.9+
4GB+ RAM
GPU recommended (optional)

Installation

# Clone repository
git clone https://github.com/yourusername/pcb-defect-detection.git
cd pcb-defect-detection

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Download DeepPCB dataset
git clone https://github.com/tangsanli5201/DeepPCB.git

Quick Start

Use Pretrained Model:

import torch
from torchvision import transforms
from PIL import Image

# Load model
model = PCBDefectClassifier()
model.load_state_dict(torch.load('models/best_model_regularized.pth'))
model.eval()

# Prepare image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

img = Image.open('path/to/pcb_image.jpg')
img_tensor = transform(img).unsqueeze(0)

# Predict
with torch.no_grad():
    predictions = model(img_tensor)

# Results
defects = ['open', 'short', 'mousebite', 'spur', 'copper', 'pin-hole']
for defect, conf in zip(defects, predictions[0]):
    if conf > 0.5:
        print(f"{defect}: {conf:.2%} confidence")

💡 Key Learnings & Insights

1. Generic AI ≠ Specialized Domains

Experiment: Tried BLIP (vision-language model) for automated defect descriptions

Result: Complete failure ❌

Input:  PCB with open circuit defect
BLIP:   "a circuit board with chip chip chip chip..."
Needed: "Open circuit detected in trace at confidence 94%"

Why it failed:

Trained on natural images (cats, dogs), not technical PCBs
No PCB-specific vocabulary
Can't recognize circuit patterns

Lesson: For technical domains, domain knowledge + simple rules > fancy AI without fine-tuning

2. Training Stability Matters

Original training showed erratic validation loss with dramatic dips. Solutions tried:

Approach	Impact
More epochs (20)	✅ Better convergence
Dropout + Weight Decay	✅✅ Significant improvement

Result: Validation loss std dev reduced by 81%

3. Default Threshold Often Optimal

Explored 1000+ threshold combinations. Finding: Default 0.5 was already near-optimal!

This indicates:

Model calibration is good
Transfer learning worked well
No need for complex threshold tuning

4. Business Context > Raw Accuracy

Question: Should we optimize for 92% F1 vs current 91.2%?

Analysis:

Development time: 8+ hours
Performance gain: ~2% F1
Business impact: 11% fewer manual reviews
Annual savings: ~$300

Decision: Not worth it. Time better spent on deployment prep.

🎯 Business Impact & ROI

Cost-Benefit Analysis

Aspect	Manual Inspection	Our AI System	Savings
Setup Cost	$0	$2,000	-
Annual Labor	$120K-300K	$10K maintenance	$110K-290K/year
Throughput	20-30 boards/hour	2000+ boards/hour	50-100× faster
Detection Rate	~85%	98.3%	Fewer field failures
Consistency	Varies	24/7 consistent	No degradation

ROI Timeline: 6-12 months

Target Industries

✅ Consumer electronics (smartphones, IoT devices)
✅ Automotive (ADAS, EV battery management)
✅ Medical devices (pacemakers, imaging equipment)
✅ Aerospace & defense (avionics, satellites)
✅ Contract manufacturers (high-volume production)

🔮 Future Enhancements

Short-Term (1-3 months)

Web Dashboard - Flask/Streamlit UI for inspectors
Defect Localization - Add bounding boxes (YOLO v8)
Confidence Calibration - Platt scaling for better probabilities

Medium-Term (3-6 months)

Active Learning - Continuously improve with production data
Explainable AI - GradCAM visualization showing detection reasons
Ensemble Models - Combine multiple architectures

Long-Term (6-12 months)

Root Cause Analysis - ML model to predict defect causes
End-to-End Platform - Integrate with MES/ERP systems
Edge Deployment - On-device inference with TensorRT/ONNX

🤔 Limitations & Considerations

Honest Assessment:

Dataset Limitations

Size: 1,500 images vs 100K+ in commercial systems
Diversity: Single PCB type; may not generalize to flex PCBs, HDI, RF boards
Class Imbalance: Real manufacturing has 100:1 defect-to-clean ratios

Architecture Limitations

No Localization: Detects presence, not exact defect location
Fixed Input Size: 640×640px; may miss small defects on large boards

Deployment Considerations

False Alarms: Some false positives (acceptable in QC)
Novel Defects: Model only knows 6 trained classes
Environmental Factors: Lighting, camera angle affect performance

📚 References & Resources

Academic Papers

Tang et al., "PCB Defects Detection Using Deep Learning", arXiv 2019
He et al., "Deep Residual Learning for Image Recognition", CVPR 2016

Datasets

DeepPCB: github.com/tangsanli5201/DeepPCB

Tools & Frameworks

PyTorch: pytorch.org
OpenCV: Image processing

🙏 Acknowledgments

DeepPCB Team - For open-sourcing the dataset
PyTorch Community - Excellent framework
ResNet Authors - Transfer learning foundation

📧 Contact

Fardin Hossain Tanmoy
🎓 MSc Data Science | New York Institute of Technology 🎓 BEng Electrical & Electronics Engineering | University of Southampton

📧 Email: fardintonu@gmail.com

🔗 LinkedIn:www.linkedin.com/in/fardin-hossain-tanmoy

📂 GitHub: https://github.com/fardinhossain007

⭐ If You Found This Useful

Give it a star ⭐ on GitHub!

Built with 🔥 by an EEE grad exploring the intersection of hardware and AI

"The best AI solution balances accuracy, speed, cost, and interpretability."

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
notebooks		notebooks
results		results
src		src
.gitignore		.gitignore
README.md		README.md
requirements.txt		requirements.txt

fardinhossain007/pcb-defect-detection

Folders and files

Latest commit

History

Repository files navigation