🔍 3002_Project - Weapon Detection with YOLOv12, ResNet Classifier, and Grad-CAM

This project detects five types of weapons in X-ray images using a combination of:

• YOLOv12 for object detection
• ResNet18 for secondary classification
• Grad-CAM for visual explainability

✅ Developed for the Deep Learning course (AIN3002)

---

📂 Dataset

• Name: SIXray
• Link: Roboflow Dataset
• Classes:
  • Gun
  • Knife
  • Pliers
  • Scissors
  • Wrench
• Custom split:
  • Train: 497 images
  • Validation: 105 images
  • Test: 105 images

---

📁 Project Structure

📦3002_Project
┣ 📁Data/
┃ ┣ 📁train/
┃ ┣ 📁val/
┃ ┗ 📁test/
┣ 📁Crops/               # YOLO-detected objects for classifier
┣ 📁results/             # mAP curves, predictions, visualizations
┣ 📄threat_classifier.pt # Trained ResNet18 model
┣ 📄best.pt              # YOLOv12 weights
┣ 📄gradcam_classifier.py
┣ 📄crop_objects.py
┣ 📄train_resnet_classifier.py
┣ 📄data.yaml
┣ 📄yolo12s.yaml
┣ 📄README.md

---

📦 Required Libraries

Install with:

pip install -r requirements.txt

Main Dependencies:

• torch
• torchvision
• ultralytics
• opencv-python
• pillow
• pytorch-grad-cam
• matplotlib

---

🚀 How to Run

1. Train YOLOv12 on a custom dataset

from ultralytics import YOLO
model = YOLO("yolo12s.yaml")
model.train(data="data.yaml", epochs=200, imgsz=1024)

2. Crop images using YOLO predictions

python crop_objects.py

3. Train ResNet classifier on cropped images

python train_resnet_classifier.py

4. Run Grad-CAM to visualize attention

python gradcam_classifier.py

---

📈 Evaluation Metrics (YOLOv12)

Class	Precision	Recall	mAP@50	mAP@50-95
Gun	0.732	0.684	0.749	0.325
Knife	0.343	0.360	0.324	0.156
Pliers	0.528	0.418	0.434	0.204
Scissors	0.445	0.267	0.257	0.119
Wrench	0.175	0.125	0.166	0.082

---

🧠 Project Highlights

• 🔍 Two-Stage Detection: YOLOv12 for detection → ResNet for refinement
• 🧠 Explainability: Grad-CAM highlights what the classifier sees
• 🧪 Strong Augmentations: Mosaic, Mixup, HSV, Flip, etc.
• 💻 Google Colab Compatible: Code tested on Google Colab Pro
• 🔐 Classifier Confidence Thresholding to reject low-confidence results

---

📌 Future Work

• Add a webcam or a real-time app
• Replace ResNet18 with EfficientNet or ViT
• Improve dataset balance (add non-weapon samples)
• Deploy via Streamlit or Flask (optional)

---

🙌 Acknowledgements

• Ultralytics YOLO
• Roboflow
• PyTorch Grad-CAM