README.md

metadata

license: mit
datasets:
  - TheDeepDas/Phytoplankton-Enumeration
language:
  - en
base_model:
  - microsoft/resnet-18
  - google/mobilenet_v2_1.0_224

🌿 Phytoplankton Detection & Automation Pipeline

🧠 Repository Overview

This repository hosts a collection of deep learning models and automation pipelines designed for phytoplankton detection and classification from microscopic or satellite-based imagery. The system is built to enable end-to-end automation — from raw image preprocessing to final ensemble-based predictions.

---

🧩 Model Collection

Model	Type	Description	Use Case
ResNet18	🧱 Base Model	Standard CNN backbone trained for robust phytoplankton feature extraction.	Baseline accuracy and interpretability.
MobileNetV2	⚡ Lightweight Model	Efficient and fast model optimized for deployment on low-resource devices.	Real-time and edge applications.
Preprocessing Module	🧼 Data Pipeline	Handles denoising, resizing, augmentation, and normalization for optimal model input.	Use before inference or training.
Full Pipeline (End-to-End)	🔄 Ensemble Model	Combines preprocessing, model inference, and ensemble aggregation for automated detection.	Plug-and-play complete solution.

---

⚙️ Pipeline Architecture

graph TD
A[Raw Input Images] --> B[Preprocessing Module 🧼]
B --> C[ResNet18 🧱]
B --> D[MobileNetV2 ⚡]
C --> E[Feature Fusion & Ensemble 🔗]
D --> E
E --> F[Phytoplankton Class Prediction 🌿]

Each stage is modular, allowing researchers and developers to integrate or replace components as needed.

---

🧪 Training Details

• Framework: PyTorch
• Optimizers: Adam / SGD
• Loss Function: CrossEntropyLoss
• Dataset: Curated Phytoplankton Image Dataset (custom/preprocessed)
• Input Size: 224×224 RGB
• Augmentations: Random rotations, flips, and color normalization
• Ensemble Strategy: Weighted average of model logits

---

📊 Performance Summary

Model	R² Score	F1 Score	Notes
ResNet18	0.897	0.93	Strong baseline with balanced precision & recall.
MobileNetV2	0.890	0.922	Lighter and faster, ideal for deployment.
Full Pipeline	0.91	0.96	Ensemble with preprocessing yields highest performance.

---

💡 Key Features

• 🧼 Automated Preprocessing — Handles normalization, resizing, and augmentation seamlessly.
• 🔗 Modular Ensemble — Combine multiple architectures dynamically.
• ⚡ Lightweight & Deployable — Optimized for both GPU and CPU inference.
• 🧬 Domain-Specific Optimization — Tuned specifically for phytoplankton morphology and color features.

---

📘 Citation

If you use this work, please cite it as:

@misc{deepdas2025phytoplankton,
  author       = {Das, Deep},
  title        = {Phytoplankton Detection and Automation Pipeline},
  year         = {2025},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/TheDeepDas/phytoplankton-Models}
}

---

❤️ Acknowledgements

Developed by Deep Das 🔗 GitHub • Hugging Face • LinkedIn

---

“Empowering ocean research through automation and intelligence.”