Mango Detection using YOLO

Overview

This project implements a robust mango detection system using various versions of the YOLO (You Only Look Once) object detection algorithm. The system is designed to identify and classify mangoes in images, providing accurate bounding box detection for agricultural applications.

Project Structure

├── Compare.ipynb          # Notebook for comparing different model versions
├── Models                 # Trained model weights
│   ├── mango_detector_v10.pt
│   ├── mango_detector_v12.pt
│   └── mango_detector_v8.pt
├── README.md
├── Report.md              # Project report in IEEE format
└── Training               # Training notebooks for different YOLO versions
    ├── v10.ipynb
    ├── v12.ipynb
    └── v8.ipynb

Features

• Multi-version YOLO Implementation: Implements YOLOv8, YOLOv10, YOLOv11, and YOLOv12 for comparative analysis
• GPU-Accelerated Training: Optimized for training with GPU support
• Comprehensive Metrics: Evaluation metrics including mAP, precision, recall, and F1-score
• Model Comparison: Tools to compare inference speed and detection accuracy across models
• Interactive Visualization: Visual representation of detection results and performance metrics

Dataset

The project uses this mango image dataset from Roboflow:

• Source: Luigui Andre Cerna Grados' "Clasificación-de-mangos" dataset
• Version: 16

Installation

Requirements

• Python 3.8+

Setup

pip install ultralytics roboflow opencv-python matplotlib pandas seaborn numpy torch

Usage

Training a Model

1. Open the desired training notebook (e.g., Training/v12.ipynb)
2. Run the notebook cells to:
   • Download and preprocess the dataset
   • Train the model with defined parameters
   • Evaluate and save the model

Comparing Models

1. Open Compare.ipynb
2. Upload the trained models from the Models directory
3. Run the comparison cells to:
   • Evaluate metrics across different model versions
   • Compare inference speed and detection accuracy
   • Visualize detection results side by side

Inference

from ultralytics import YOLO

# Load model
model = YOLO('Models/mango_detector_v12.pt')

# Run inference
results = model.predict(
    source='path/to/image.jpg',
    conf=0.5,
    save=True
)

# Process results
for r in results:
    boxes = r.boxes
    print(f"Found {len(boxes)} mangoes in the image")

Model Performance

A comparison of the model versions:

Model	mAP@0.5	mAP@0.5:0.95	Precision	Recall	F1-Score	Inference Time (s)	Training Time
YOLO v8	0.9087	0.6425	0.9052	0.8534	0.8785	13.3864	49 min
YOLO v10	0.8971	0.6178	0.8900	0.8305	0.8592	7.4955	55.5 min
YOLO v11	0.9194	0.6410	0.9341	0.8336	0.8810	7.0858	52.2 min
YOLO v12	0.9162	0.6412	0.9165	0.8291	0.8706	7.4938	69.3 min

Development Process

1. Data Preprocessing: Images are normalized and augmented to enhance training
2. Model Training: Models are trained with optimized hyperparameters
3. Evaluation: Comprehensive metrics are calculated to assess model performance
4. Comparison: Different YOLO versions are compared to identify the best-performing model

License

This project is licensed under the terms of the MIT license.

Acknowledgments

• Dataset provided by Luigui Andre Cerna Grados
• Built with Ultralytics YOLO implementation
• Developed using Google Colab for GPU acceleration