YOLOv8-nano model trained on synthetic holographic microplastic dataset. Contributes to SDG Goal 6: Clean Water and Sanitation.
Complete Setup Guide: See docs/setup.md for detailed step-by-step instructions.
Note: The repository includes a pre-trained model. You can run inference immediately without downloading datasets. Datasets are only needed for retraining.
- Open
colab_train.ipynb - Upload to Google Colab
- Runtime → Change runtime type → T4 GPU
- Run all cells
Training Time: ~2-3 hours on T4 GPU (50 epochs)
pip install ultralytics opencv-python pandas tqdm
# Download datasets (see Dataset Download section below)
# Extract HMPD-Gen.zip and microplastic_data.zip
# Generate dataset
python 01_generate_synthetic_dataset.py
# Train model
python 02_train_yolov8_colab.pypip install ultralytics opencv-python flask
# Use pre-trained model for detection
python 03_esp32_integration.py --esp32 192.168.4.1 --conf 0.35├── docs/
│ └── setup.md # Complete setup guide for beginners
├── esp32_code/ # ESP32-CAM Arduino code
│ ├── microplastic_detector.ino # Main sketch
│ ├── app_httpd.cpp # HTTP streaming server
│ ├── app_httpd.h # Server header
│ └── camera_pins.h # Pin definitions
├── colab_train.ipynb # Google Colab training notebook
├── 01_generate_synthetic_dataset.py # Dataset generation script
├── 02_train_yolov8_colab.py # Training script
├── 03_esp32_integration.py # Real-time detection with ESP32-CAM
├── yolov8_microplastic_trained.pt # Trained model weights (3.2M params, 6MB)
├── confusion_matrix.png # Classification accuracy matrix
├── results.png # Training metrics graphs (loss, mAP, precision, recall)
├── test_results/ # 20 sample detection images with predictions
│ ├── result_01_microplastic_synthetic_00817.jpg
│ ├── result_02_microplastic_synthetic_01500.jpg
│ └── ... (18 more examples)
├── README.md # This file
├── LICENSE # MIT License
└── CITATION.cff # Citation metadata
Training Data: 15,106 holographic patches from HMPD dataset (both gt.csv and gtPossible.csv)
Synthetic Dataset: 2,000 images (640×640) with 1-8 particles per scene
Source: HMPD Repository by CNR-ISASI
Paper: Cacace, T., Del-Coco, M., Carcagnì, P., Cocca, M., Paturzo, M., & Distante, C. (2023). "HMPD: A Novel Dataset for Microplastics Classification with Digital Holography." In Image Analysis and Processing – ICIAP 2023 (pp. 123-133). Springer-Verlag. https://doi.org/10.1007/978-3-031-43153-1_11
The training dataset files are hosted in GitHub Releases:
Download the dataset files:
- HMPD-Gen.zip (151 MB) - Holographic microplastic dataset
- microplastic_data.zip (161 MB) - Processed synthetic dataset
Extract the datasets:
# Download (use browser or wget/curl)
wget https://github.com/niloydebbarma-code/microplastic-detection/releases/download/v1.0.0/HMPD-Gen.zip
wget https://github.com/niloydebbarma-code/microplastic-detection/releases/download/v1.0.0/microplastic_data.zip
# Extract in project root
unzip HMPD-Gen.zip
unzip microplastic_data.zip
# Or on Windows (PowerShell)
Expand-Archive HMPD-Gen.zip -DestinationPath .
Expand-Archive microplastic_data.zip -DestinationPath .Note: The pre-trained model yolov8_microplastic_trained.pt is included in the repository, so you can run inference immediately without downloading datasets.
- Architecture: YOLOv8-nano (3.2M parameters)
- Training: 50 epochs on T4 GPU (~2-3 hours)
- Detection Range: 10-500 μm microplastic particles
- Input Size: 640×640 pixels
Training Validation: conf=0.25 (default for model evaluation)
- Used during training to measure model performance
- Lower threshold captures more detections for comprehensive metrics
- Evaluates model's ability to detect particles at various confidence levels
Real-time Inference: Configurable via --conf parameter
- Default:
conf=0.25(balanced - more detections, may include false positives) - Recommended for production:
conf=0.35(balanced precision and detection rate) - Optimized threshold reduces false alarms while maintaining good detection
# Use optimal confidence for accurate real-time detection
python 03_esp32_integration.py --esp32 192.168.4.1 --conf 0.35results.png - Training metrics over 50 epochs:
- Loss curves: Training/validation loss (box, class, DFL losses)
- mAP50: Mean Average Precision at IoU threshold 0.5
- mAP50-95: mAP averaged across IoU thresholds 0.5 to 0.95
- Precision: Ratio of true positives to all positive predictions
- Recall: Ratio of detected particles to all ground truth particles
Training and validation metrics showing model convergence over 50 epochs
confusion_matrix.png - Classification accuracy analysis:
- Shows true positives, false positives, false negatives for microplastic detection
- Diagonal values indicate correct detections
- Off-diagonal values show misclassifications
- Background class performance also included
Classification accuracy matrix showing model detection performance
The test_results/ folder contains 20 sample detection images showing:
- Bounding boxes around detected microplastic particles
- Confidence scores for each detection (calibrated to 0.35 threshold)
- Particle size estimates (10-500 μm range)
- Detection performance on synthetic holographic images
Example files:
result_01_microplastic_synthetic_00817.jpgthroughresult_20_microplastic_synthetic_01666.jpg- Each shows model predictions overlaid on test images
- Demonstrates detection accuracy across different particle counts and sizes
Sample detection results showing bounding boxes and confidence scores
- Training Set: 14,000 images from 15,106 HMPD holographic patches
- Validation Set: 1,106 images for model evaluation
- Synthetic Data: 2,000 generated images (640×640) with 1-8 particles per scene
- Augmentation: Rotation, brightness, contrast variations applied during training
If you use this work, please cite:
@software{microplastic_detection_2026,
title = {Automated Microplastic Detection and Quantification using Deep Learning},
author = {Niloy Deb Barma and Godugula Tejasi},
year = {2026},
publisher = {GitHub},
url = {https://github.com/niloydebbarma-code/microplastic-detection}
}- Niloy Deb Barma (niloydebbarma-code)
- Godugula Tejasi (godugulatejaswi)
- Dr. P. Sathyaseelan, Associate Professor
- Dr. M. R. Arun, Professor
Dataset Citation:
@inproceedings{10.1007/978-3-031-43153-1_11,
author = {Cacace, Teresa and Del-Coco, Marco and Carcagn\`{\i}, Pierluigi and Cocca, Mariacristina and Paturzo, Melania and Distante, Cosimo},
title = {HMPD: A Novel Dataset for Microplastics Classification with Digital Holography},
year = {2023},
isbn = {978-3-031-43152-4},
publisher = {Springer-Verlag},
address = {Berlin, Heidelberg},
url = {https://doi.org/10.1007/978-3-031-43153-1_11},
doi = {10.1007/978-3-031-43153-1_11},
booktitle = {Image Analysis and Processing – ICIAP 2023: 22nd International Conference, ICIAP 2023, Udine, Italy, September 11–15, 2023, Proceedings, Part II},
pages = {123–133},
numpages = {11},
keywords = {dataset, microplastic, holography, deep learning},
location = {Udine, Italy}
}- HMPD Dataset: CNR-ISASI for holographic microplastic dataset
- Ultralytics: YOLOv8 framework
- SDG 6.3: Clean water and sanitation
- Component/financial support (non-coding): S. Yoga Lakshmi, Chavali Pavan Siva Sri Balaji, L. Santhosh
MIT License - See LICENSE file for details.
Found a bug or want to contribute? Open an issue or pull request!
For questions or collaborations, please open an issue on GitHub.
Keywords: Microplastic Detection, Digital In-Line Holography, Lensless Imaging, YOLOv8, ESP32-CAM, Water Quality Monitoring, Frugal Science, SDG 6.3, Computer Vision, Environmental