martinbadrous/Microparticle-Detection-YOLOv5

🧫 High-Speed Microparticle Detection and Tracking using YOLOv5

Author: Martin Badrous
Affiliation: ViBOT Master’s Program, Bourgogne University
Year: 2021
License: MIT

---

🧠 Abstract

This repository hosts the deep learning model developed for the master’s thesis
“High-Speed Microparticle Detection and Tracking” (2021),
conducted under the ViBOT Master’s Program, Bourgogne University.

The model implements a YOLOv5-based object detection pipeline for identifying and counting PMMA microplastic particles in high-speed microscopy imagery.
It represents the second and deep learning phase of the research, following an earlier classical image processing approach.

---

📚 Research Context

Microparticle detection is a key task in studying particulate contamination and flow behavior under a microscope.
Traditional methods based on thresholding and contour detection struggle with illumination, blur, and focus variations.
To overcome these limitations, this work explores a deep learning detector (YOLOv5) trained to recognize two microplastic classes under 20× magnification.

Target Classes

• PMMA10
• PMMA20

Model Architecture

• Model: YOLOv5s (PyTorch)
• Input Resolution: 640 × 640
• IoU Threshold: 0.5
• Confidence Threshold: 0.5
• Min Detection Area: 500 px
• Scale: 0.30 µm²/pixel
• Epochs: 50
• Batch Size: 16

---

🧪 Dataset and Training

• Dataset Split: 70% train / 20% validation / 10% test
• Data Source: Custom microscope videos at 20× magnification
• Augmentations: Gaussian noise, blur, brightness, contrast, and saturation (rotation disabled)
• Optimization: SGD with cosine learning rate decay
• Framework: PyTorch (YOLOv5 Ultralytics)

Evaluation Results

Metric	PMMA10	PMMA20	Mean
Precision	0.80	0.79	0.795
Recall	0.88	0.91	0.895
mAP@0.5	0.97	0.97	0.97

---

🧩 Post-Processing and Analysis

After detection, each particle undergoes morphological analysis via three computed features:

Feature	Description
Otsu Binarized Area (px)	Measures the segmented particle area
Laplacian Variance	Quantifies image focus/sharpness
Chi-Square Histogram Distance	Measures similarity to a reference “ideal” particle histogram

These features are visualized in a 3D scatter space to assess detection focus, size distribution, and classification consistency.

---

🧭 Model Structure

Microparticle-Detection-YOLOv5-HF/
├── README.md          # Academic model card (this file)
├── config.yaml        # Metadata (architecture and params)
├── model_card.json    # Structured metadata for Hugging Face Hub
└── best.pt            # YOLOv5 trained weights

---

🧩 How to Use

import torch

# Load custom YOLOv5 model
model = torch.hub.load('ultralytics/yolov5', 'custom', path='best.pt')
results = model('microscope_sample.jpg')
results.show()

---

📘 Citation

Badrous, M. (2021). High-Speed Microparticle Detection and Tracking.
ViBOT Master’s Program, Bourgogne University.

---

🧑‍💻 Author

Martin Badrous
ViBOT Master’s Program, Bourgogne University (2021)

---

🪶 License

This model is released under the MIT License.

---

🧭 Future Work

• Integrate motion tracking and particle trajectory reconstruction
• Extend dataset to diverse microplastic materials and shapes
• Deploy as an interactive Hugging Face Space (Streamlit demo)
• Explore YOLOv8 or RT-DETR for real-time microscopy inference

---