🌿 Semi-Supervised Weed Detection using YOLOv9

A semi-supervised deep learning pipeline for real-time weed detection in precision agriculture using YOLOv9, GIoU loss, and pseudo-labeling — achieving high accuracy with only 200 labeled images.

---

📁 Project Structure

📦 Semi_Supervise_Weed_Detection
 ┣ 📓 01_EDA_Augmentation.ipynb       # EDA and augmentation pipeline
 ┣ 📓 02_Train_Inference.ipynb        # Model training and inference
 ┣ 📄 Semi_Supervised_Weed_Detection_Report.pdf
 ┣ 📄 requirements.txt
 ┗ 📄 README.md

---

📊 Results

Metric	Baseline YOLOv9	Final Model
Precision	0.77	0.89
Recall	0.85	0.88
F1 Score	0.81	0.89
mAP@50-95	0.58	0.62
Final Metric	0.69	0.75

---

🔍 Methodology

1. Dataset

• 200 labeled images annotated with crops and weeds
• 1,000 unlabeled images used for semi-supervised training

2. Data Augmentation

Applied using Albumentations — each image augmented 5x to simulate real-world conditions:

• RandomBrightnessContrast, HueSaturationValue
• RandomFog, RandomRain, RandomSnow
• MotionBlur, GaussianBlur, GaussianNoise
• CLAHE, GridDistortion

3. Semi-Supervised Learning Pipeline

1. Initial Training — YOLOv9 trained on 200 labeled images
2. Pseudo-Labeling — model predicts labels for 1,000 unlabeled images
3. Confidence Filtering — top 200 predictions with confidence ≥ 0.5 selected
4. Iterative Retraining — model retrained on expanded dataset

4. Loss Functions

• GIoU Loss — improved bounding box alignment under complex field conditions
• F1 Score Loss — balances precision and recall to minimize false positives and negatives

---

🚀 Getting Started

1. Clone the repo

git clone https://github.com/Shradd7/Semi_Supervise_Weed_Detection.git
cd Semi_Supervise_Weed_Detection

2. Install dependencies

pip install -r requirements.txt

3. Download model weights

Pre-trained weights available here → Google Drive

Place the weights file in the root directory.

4. Run notebooks in order

01_EDA_Augmentation.ipynb    — run first
02_Train_Inference.ipynb     — run second

---

🛠️ Tech Stack

Category	Tools
Model	YOLOv9 (Ultralytics)
Framework	PyTorch
Augmentation	Albumentations
Image Processing	OpenCV
Visualization	Matplotlib, Seaborn
Data	NumPy, Pandas

---

🔮 Future Work

• Deploy as a real-time inference API using FastAPI
• Extend to multi-class weed species detection
• Test on drone footage for field-scale deployment
• Experiment with FixMatch or MixMatch for stronger semi-supervised baselines

---

👥 Authors

• Shradd7
• Abhiraj Mandal
• Abhi Tanwar
• Arnav Gupta
• Tushay Gupta
• Mit Narodia