Edge Vision AI

Edge-Based Wafer Defect Detection Intelligence

Edge Vision AI is an end-to-end artificial intelligence system for automated wafer defect detection, explainability, and deployment on NXP edge devices. The project is designed for real-time, low-power, and scalable semiconductor quality inspection.

Project Resources

Dataset

Wafer Inspection Pipeline

🎥 Demo Video

Click here to watch the project demo

Problem Statement

Manual wafer inspection in semiconductor manufacturing is slow, costly, error‑prone, and difficult to scale for high‑volume production. Existing inspection systems lack real‑time intelligence and flexibility, creating a need for a lightweight, automated, and reliable solution.

Proposed Solution

Edge Vision AI provides an intelligent inspection pipeline that:

• Automatically classifies wafer defects
• Runs efficiently on edge hardware
• Explains model predictions
• Learns from uncertain samples
• Operates in offline environments

---

🚀 Performance Snapshot(Key Highlights)

Metric	Result
Model Size	1.9 MB
Accuracy	75%+ overall
Training Accuracy	~69%
Validation Accuracy	~79%
Inference Speed	Real-time
Memory Usage	Low
Deployment	Edge-ready (NXP compatible)
Training Time	~5 mins

---

Key Innovations

• Lightweight Edge AI

  • Compact, embedded‑ready CNN
  • Low‑latency, real‑time inference

• Self‑Learning System

  • Automatically stores low‑confidence cases
  • Model keeps improving with new data

• Edge Deployment

  • INT8 quantized model
  • Runs on NXP hardware with eIQ support

• Smart Analytics

  • Batch inspection + yield calculation
  • CSV logging with offline support

Model Accuracy and Validation Performance

• Training Accuracy: ~69%
• Validation Accuracy: ~79%

System Overview

The system follows a modular pipeline:

Image Acquisition → Preprocessing → AI Inference → Confidence Evaluation
                                           ↓
                                Explainable AI Visualization
                                           ↓
                               Edge Deployment and Monitoring

Training and deployment pipelines are separated for optimization and reliability.

---

Dataset File Structure

Dataset/
│
├── Train/
│   ├── bridge/
│   ├── clean/
│   ├── cmp/
│   ├── crack/
│   ├── ler/
│   ├── open/
│   ├── others/
│   └── vias/
│
├── Validation/
│   ├── bridge/
│   ├── clean/
│   ├── cmp/
│   ├── crack/
│   ├── ler/
│   ├── open/
│   ├── others/
│   └── vias/
│
└── Self-learning/
    └── low_confidence_samples/

---

Technology Stack

• Machine Learning: TensorFlow, Keras, NumPy
• Image Processing: OpenCV
• Data Management: Pandas, CSV
• Deployment: TensorFlow Lite, NXP eIQ Toolkit, C/C++
• Development Tools: Python, Google Colab, GitHub

---

Project Structure

---

System Workflow

Step 1: Data Input

Wafer images are captured using camera modules or uploaded manually.

Step 2: Preprocessing

Images are resized, normalized, and formatted for model input.

Step 3: Inference

The CNN model predicts defect class and confidence score.

Step 4: Confidence Evaluation

• High confidence → Accepted result
• Low confidence → Stored for retraining

Step 5: Explainability

Grad-CAM generates visual explanations highlighting defect regions.

Step 6: Edge Deployment

Optimized model runs on NXP hardware for real-time inspection.

---

Interpretation

The close alignment between training and validation accuracy indicates:

• Minimal overfitting
• Stable learning behavior
• Good generalization performance

This balance confirms that the model performs consistently on both known and unknown wafer images.

Impact

Industrial Impact

• Reduced inspection costs
• Improved manufacturing yield
• Faster quality validation

Technical Impact

• Promotes Edge AI adoption
• Improves model transparency
• Enables scalable deployment

Innovation Impact

• Combines XAI and embedded AI
• Supports adaptive learning
• Bridges research and production

---

Hackathon Highlights

• Complete end-to-end solution
• Edge-ready deployment
• Explainable decision system
• Self-learning pipeline
• Industry-focused design

This project demonstrates the transition from prototype to deployable system.

---

Future Enhancements

• Automated retraining pipeline
• Cloud-based monitoring dashboard
• Live camera integration
• Multi-device deployment
• Predictive maintenance features

---

Team

• Khusi Mittal
• Mithun Srinivas

---

Definition

Training Accuracy represents how well the model learns patterns from the training dataset.
It indicates the model’s ability to fit known data. Validation Accuracy represents how well the trained model performs on unseen data.
It measures the model’s generalization capability in real-world scenarios.

---

Acknowledgements

• Hackathon Organizers
• NXP eIQ Platform
• Open Source Community

Thank you for reviewing WaferGuard AI.