Underwater Microplastic Detection System

An AI-powered system that classifies underwater microplastic particles from microscopy images using deep learning (EfficientNet-B2). Upload an image to the live web dashboard and get an instant contamination state with per-class confidence scores.

---

Dashboard

---

Prediction Example

---

Features

• 5-class polymer classification — Alkyd, Bead, Cellophane, Degraded, Fiber
• 5 contamination severity levels (State 0 Clean → State 4 Critical)
• Real-time web dashboard with animated underwater UI
• Drag-and-drop image upload with instant results
• Prediction history log (last 8 analyses)
• Per-class confidence bar chart
• 97.3% validation accuracy on held-out test set
• Runs on CPU — no GPU required

---

Project Structure

Underwater_Microplastic_Detection/
├── train.py          # Model training (EfficientNet-B2 + CosineAnnealing)
├── verify.py         # Quick accuracy check on 10 random validation samples
├── dashboard.py      # Flask web app — live detection interface
├── dataset/
│   ├── data.yaml     # Dataset config (paths, class names)
│   ├── train/
│   │   ├── images/   # Training images (.jpg)
│   │   └── labels/   # YOLO-format label files (.txt)
│   └── valid/
│       ├── images/   # Validation images (.jpg)
│       └── labels/
└── visuals/
    ├── Dashboard.png
    └── predic.png

---

Setup

1. Install dependencies

pip install torch torchvision flask scikit-learn seaborn matplotlib pillow

2. Prepare dataset

Place the dataset under C:\microplastic\dataset\ with the following layout:

C:\microplastic\dataset\
├── train\
│   ├── images\   ← training .jpg files
│   └── labels\   ← YOLO .txt label files
└── valid\
    ├── images\
    └── labels\

3. Train the model

python train.py

Trains for 60 epochs with CosineAnnealingLR. Saves the best checkpoint to C:\microplastic\models\best_model.pth and plots to C:\microplastic\results\.

4. (Optional) Verify predictions

python verify.py

Runs 10 random validation images through the saved model and prints actual vs predicted class with confidence.

5. Launch the dashboard

python dashboard.py

Open http://127.0.0.1:5001 in your browser.

---

Model Details

Property	Value
Architecture	EfficientNet-B2
Input size	260 × 260 px
Pretrained on	ImageNet
Classes	5 (alkyd, bead, cellophane, degraded, fiber)
Training images	453
Validation images	204
Epochs	60
Optimizer	AdamW (lr=3e-4, weight_decay=1e-4)
Scheduler	CosineAnnealingLR
Loss	CrossEntropy + label smoothing 0.05
Best val accuracy	97.3%
Device	CPU / GPU (auto-detected)

---

Classes & Contamination States

Class	State	Severity	Hazard
Alkyd	1 — Trace debris	Low	Low
Bead	2 — Moderate contamination	Medium	Medium
Cellophane	2 — Moderate contamination	Medium	Medium
Degraded	3 — Heavy contamination	High	High
Fiber	4 — Critical pollution	Critical	Danger

---

Tech Stack

• Python 3.10
• PyTorch + TorchVision (EfficientNet-B2)
• Flask (web dashboard)
• Scikit-learn (metrics)
• Matplotlib + Seaborn (training plots)
• Pillow (image loading)