Wafer Defect Detection - Lightweight Mask R-CNN

🧪 Overview

This project aims to build a lightweight and efficient Mask R-CNN-based model to detect single and mixed-type defects in wafer maps.
The model incorporates:

• MobileNetV3 + CBAM as the backbone,
• FCOS (Anchor-Free RPN) for object proposal,
• GhostConv for efficient mask head processing.

The goal is to achieve real-time defect detection suitable for semiconductor production environments.

---

📁 Dataset

• WM-811K: Kaggle WM-811K Dataset
• Mixed-Type WM38: Kaggle Mixed-Type Dataset

We apply preprocessing including:

• Noise removal, resizing, normalization
• Data augmentation (rotation, flipping) to address class imbalance

---

🧠 Model Architecture

• MobileNetV3 + CBAM
• FCOS
• RoI Align + FC layers
• Classification + BBox Regression
• GhostConv-based Mask Head

⚡ Designed for efficient instance segmentation and pixel-wise defect localization.

---

📊 Evaluation Metrics

We use both detection and efficiency metrics:

Detection:

• Precision / Recall / F1-Score
• Accuracy
• IoU
• Multi-label Classification Accuracy

Efficiency:

• FLOPs
• Model Size (MB)
• Parameter Count
• FPS (Frames Per Second)

---

🔧 Tools

• Python, PyTorch
• Google Colab Pro
• OpenCV, Matplotlib
• Scikit-learn, NumPy

---

👨‍💻 Team Members

Name	Student ID
백지원	2033014
정윤수	2172045
한지민	2172126
황수빈	2176422
김지우	2272008

---

📅 Timeline

• Week 1–3: Project planning, goal setting
• Week 2–6: Dataset collection and preprocessing
• Week 4–11: Model development, tuning, and evaluation
• Regular team meetings: 3 times/week

---

📌 Keywords

Semiconductor, Wafer Defect, Instance Segmentation, Mask R-CNN, MobileNetV3, GhostConv, FCOS, Edge AI, Lightweight Model

---

🔗 References

• GhostNet (CVPR 2020)
• MobileNetV3 (ICCV 2019)
• CBAM (ECCV 2018)
• FCOS (ICCV 2019)
• WM-811K & Mixed-Type Wafer Datasets on Kaggle