gdhanush270/3d-model-defect-identifier

Automated Defect Detection in 3D Mesh Files Using Multi-Model Deep Learning Approaches

📌 Project Overview

This project introduces a multi-modal deep learning approach to detect defects in 3D mesh files by combining:

• CNN (Convolutional Neural Network) for object classification using ModelNet40 dataset images.
• GNN (Graph Neural Network) for defect identification in OFF files (3D mesh models).
• Fusion Model integrating CNN and GNN for improved classification accuracy.

📁 Dataset & Novelty

The dataset used in this project is novel and proprietary, focusing on defect detection in 3D mesh files. Only the ModelNet40 dataset is publicly available.

🔹 Folder Structure

📦 Dataset
 ┣ 📂 Images
 ┃ ┣ 📂 train
 ┃ ┃ ┣ 📂 category_1
 ┃ ┃ ┣ 📂 category_2
 ┃ ┃ ┗ ...
 ┃ ┗ 📂 test
 ┗ 📂 OFF_files
   ┣ 📂 train
   ┃ ┣ 📂 category_1
   ┃ ┃ ┣ 📂 normal
   ┃ ┃ ┗ 📂 defected
   ┃ ┣ 📂 category_2
   ┃ ┃ ┣ 📂 normal
   ┃ ┃ ┗ 📂 defected
   ┗ 📂 test
 

• Images Folder → Contains object images categorized into different classes (used for CNN).
• OFF Files Folder → Each category has "normal" and "defected" OFF files (used for GNN).

---

🚀 Model Architecture

🔹 CNN Model (Image Classification)

• Uses a pretrained CNN model (ResNet) to classify objects.

🔹 GNN Model (Defect Identification)

• Processes OFF files using node features and adjacency matrices.
• Uses a 13-layer deep GNN model to capture mesh structure defects.

🔹 Multi-Modal Fusion Model

• Combines CNN and GNN outputs using fully connected layers.
• Improves accuracy by leveraging both image and graph information.

---

⚙️ Installation & Setup

🔹 1️⃣ Install Dependencies

pip install tensorflow numpy networkx trimesh

🔹 2️⃣ Run Training

python Utils/train.py

🔹 3️⃣ Evaluate Model

python Utils/evaluate.py

---

📊 Results & Evaluation

• CNN Classification Accuracy: 76%
• GNN Defect Detection Accuracy: 78%
• Fusion Model Accuracy: 85%

---

🛠️ Future Improvements

• Use a more complex GNN model (with at least 13 layers).
• Improve multi-modal fusion model by adding extra layers.
• Train on a larger dataset to improve generalization.

---

👨‍💻 Author

Dhanush
📧 Contact: e-mail

---