Trash vs. Marine Life Classifier

A deep learning image classifier that distinguishes between ocean trash/debris and marine life, built to support automated ocean pollution detection systems.

🔗 Live Demo: Coming soon

---

Motivation

Marine debris is one of the most pressing environmental challenges facing ocean ecosystems today. Manual identification and removal of underwater trash is time-consuming and expensive. This project explores whether a convolutional neural network can reliably distinguish between ocean trash and marine life — a foundational step toward building automated debris detection and collection systems.

---

Dataset

The model was trained on two combined datasets:

Sea Animals

• Source: Sea Animals Dataset (Kaggle)
• Size: ~19,000 images
• Classes: Various marine life (fish, jellyfish, coral, etc.)

Ocean Trash / Marine Debris

• Source 1: Original project dataset
• Source 2: UW Garbage Debris Dataset (Kaggle) — 5,130 underwater debris images

Final Split

Split	Sea Animals	Trash	Total
Train	15,214	5,575	20,789
Val	3,804	1,394	5,198
Total	19,018	6,969	25,987

Note on class imbalance: The dataset has an approximate 3:1 ratio of sea animals to trash. This was a deliberate improvement from the original 10:1 ratio, achieved by adding the UW Garbage Debris dataset. Further balancing is noted as a future improvement.

---

Model

Architecture

• Base model: ResNet-18 (pretrained on ImageNet)
• Technique: Transfer learning — all layers fine-tuned, final fully connected layer replaced with a 2-class output layer
• Framework: PyTorch

Training Configuration

Parameter	Value
Optimizer	SGD (momentum=0.9)
Learning rate	0.001
LR scheduler	StepLR (step=7, gamma=0.1)
Batch size	4
Epochs	10
Hardware	NVIDIA T4 GPU (Google Colab)

Why Transfer Learning?

ResNet-18 was pretrained on ImageNet's 1.2 million images, giving it strong low-level feature representations (edges, textures, shapes). Fine-tuning this model on our domain-specific dataset allows us to leverage that prior knowledge while adapting to the specific visual characteristics of underwater imagery.

---

Results

⚠️ Results below are from the initial training run (before dataset balancing). Updated metrics will be added after retraining.

Classification Report

Class	Precision	Recall	F1-Score	Support
Sea Animals	0.90	0.45	0.60	3,804
Trash	0.08	0.48	0.13	368
Accuracy			0.46	4,172

Confusion Matrix

[Insert confusion matrix image here]

---

Error Analysis

The initial model showed signs of significant class bias due to the 10:1 imbalance in the training data:

• High trash recall (0.48) but very low precision (0.08): The model was triggering "trash" predictions too broadly, misclassifying many sea animals as trash.
• Low sea animal recall (0.45): Despite having far more sea animal training examples, the model still struggled — suggesting it defaulted to predicting trash in uncertain cases.
• Overall accuracy of 46% is close to random chance, indicating the model had not learned meaningful discriminating features and was likely exploiting class frequency patterns rather than visual content.

Root Cause

The primary issue was severe class imbalance — approximately 10 sea animal images for every 1 trash image. The model learned that predicting "sea animal" was almost always correct, leading to poor generalization on the minority class.

Fix Applied

Added ~5,130 underwater debris images from the UW Garbage Debris dataset, reducing the imbalance from 10:1 to approximately 3:1. The model was then retrained from scratch.

---

How to Run

Requirements

torch
torchvision
numpy
matplotlib
seaborn
scikit-learn

Training

1. Clone this repository
2. Organize your dataset in the following structure:

datasets/
├── train/
│   ├── sea_animals/
│   └── trash/
└── val/
    ├── sea_animals/
    └── trash/

3. Open ocean_life_vs_trash.ipynb in Google Colab
4. Mount your Google Drive and update data_dir to your dataset path
5. Run all cells in order

Inference

import torch
from torchvision import models, transforms
from PIL import Image

# Load model
model = models.resnet18()
model.fc = torch.nn.Linear(model.fc.in_features, 2)
model.load_state_dict(torch.load('model.pt'))
model.eval()

# Preprocess image
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

img = Image.open('your_image.jpg')
img_tensor = transform(img).unsqueeze(0)

# Predict
with torch.no_grad():
    output = model(img_tensor)
    _, pred = torch.max(output, 1)
    classes = ['sea_animals', 'trash']
    print(f"Prediction: {classes[pred.item()]}")

---

Future Improvements

• Further dataset balancing: Increase trash images to match sea animal count (~19,000) for a 1:1 ratio
• Data augmentation: Apply stronger augmentation on the minority class (rotation, color jitter, flipping) to synthetically balance the dataset
• Weighted loss function: Use CrossEntropyLoss with class weights to penalize misclassification of the minority class more heavily
• Larger architecture: Experiment with ResNet-50 or EfficientNet for improved feature extraction
• Deploy live demo: Build and deploy a Hugging Face Spaces demo for real-time image classification

---

Project Structure

├── ocean_life_vs_trash.ipynb   # Main training notebook
├── model.pt                    # Trained model weights
└── README.md                   # This file

---

Author

Maayan Matsliah — Computer Science @ Northeastern University, AI Concentration