TrashNet Image Classification

An image classification project using transfer learning with ResNet50 to classify waste materials into 6 categories for improved recycling and waste management.

📋 Overview

Waste contamination in recycling is a growing problem due to lack of awareness about which items are recyclable. This project builds a deep learning classifier that takes RGB images of waste materials and categorizes them into 6 classes, helping to reduce contamination and improve recycling efficiency.

🎯 Problem Statement

Improper sorting of recyclables leads to entire batches being discarded, exacerbating pollution and resource waste. This classifier aims to automate waste categorization using computer vision and deep learning.

📊 Dataset

TrashNet Dataset by Gary Thung and Mindy Yang

• Total Images: ~2,500 RGB images
• Classes: 6 categories
  • Cardboard
  • Glass
  • Metal
  • Paper
  • Plastic
  • Trash

Key Characteristics:

• Images organized in class-labeled folders
• Class imbalance present (Trash: 137 samples, Paper: 594 samples)
• Real-world waste material images with varying backgrounds

Dataset Split: 70% Train / 15% Validation / 15% Test (Stratified)

🏗️ Architecture

Transfer Learning with ResNet50

Base Model: ResNet50 pretrained on ImageNet

• Total Parameters: ~25M
• Architecture: 50-layer deep residual network
• Custom Classifier Head:

  Linear(2048 → 512) → ReLU → Dropout(0.5) → Linear(512 → 6)
  

Three Model Variants Explored

1. Model 1 (Frozen Features): Only classifier trained

   • Test Accuracy: 83.95%

2. Model 2 (Full Fine-Tuning): All layers trainable

   • Test Accuracy: 75.26%

3. Model 3 (Selective Fine-Tuning): Last conv block (layer4) + classifier

   • Test Accuracy: 91.05% ✅ Best Baseline

🔧 Data Preprocessing

Image Transformations

• Resize: 224×224 pixels
• Normalization: ImageNet mean/std
• Augmentation Pipeline:
  • Random horizontal flip
  • Random rotation (±5-10°)
  • Color jitter (brightness, contrast, saturation)
  • Random affine transformations
  • Random erasing (Aug2)
  • Gaussian blur (Aug3)

Class Balancing Strategy

• Oversampling minority classes to match maximum class count
• Augmentation applied only to oversampled images
• Result: 416 samples per class in balanced training set

📈 Results

Best Model Performance

Configuration	Test Accuracy	Notes
Baseline (No Aug, No Balance)	91.05%	Model 3 architecture
Aug2 + Balanced	92.11%	✅ Best overall
Aug1 + Balanced	89.74%	Lighter augmentation
Aug3 + Balanced	90.00%	Aggressive transforms

Augmentation Comparison

Aug1 (Geometric + Color):

• Random flip, rotation, color jitter, affine, perspective

Aug2 (Spatial + Erasing):

• Random crop, rotation, affine, random erasing
• Best performer at 92.11%

Aug3 (Filter-based):

• Gaussian blur, invert, posterize, grayscale

Key Findings

✅ Balancing + Augmentation: +1.06% improvement over baseline
✅ Selective Fine-Tuning: Better than full fine-tuning (91% vs 75%)
✅ Data Efficiency: 2,500 images sufficient with proper augmentation
✅ Cross-Domain Test: 81% accuracy on completely new waste images

💡 Data Analysis Insights

Correlation Heatmap (ResNet18 Features)

• Glass ↔ Plastic: Highest similarity (0.95+)
• Metal ↔ Glass: Strong correlation due to shine/reflectivity
• Paper ↔ Cardboard: Similar texture patterns
• Trash: Dispersed across all categories

t-SNE Clustering

• Moderately distinct clusters for each material
• Overlap between visually similar pairs (Glass-Plastic, Paper-Cardboard)
• Trash class dispersed throughout feature space

Common Misclassifications

Based on test set errors:

• Paper ↔ Cardboard (texture similarity)
• Glass ↔ Plastic (transparency/reflectivity)
• Metal ↔ Glass (surface properties)

🚀 Getting Started

Prerequisites

pip install torch torchvision numpy matplotlib scikit-learn tqdm seaborn

Installation

git clone https://github.com/VK4041/TrashNet_Image_Classification.git
cd TrashNet_Image_Classification

Dataset Setup

1. Download TrashNet dataset from GitHub
2. Place in Data/TrashNet/trashnet/ directory
3. Structure should be:

   trashnet/
   ├── cardboard/
   ├── glass/
   ├── metal/
   ├── paper/
   ├── plastic/
   └── trash/
   

Training

# Load best model configuration
model = ResNet50WasteClassifier(num_classes=6, freeze_features=True)
for param in model.base_model.layer4.parameters():
    param.requires_grad = True

# Train with balanced augmented data
train_model(model, balanced_train_loader, val_loader, test_loader, epochs=20)

🔬 Profiling & Optimization

Performance Bottleneck Analysis

Before Optimization:

• DataLoader CPU time: ~730ms
• Main bottleneck: Data loading

After Optimization (num_workers=4, pin_memory=True):

• DataLoader CPU time: ~72ms
• 90% reduction in data loading time

Training Efficiency

• Epochs: 20 with early stopping (patience=3)
• Batch size: 128
• Optimizer: Adam (lr=0.001)
• Loss: CrossEntropyLoss
• Device: CUDA (T4 GPU)

📊 Visualization Features

• Class distribution plots
• Random sample display with denormalization
• Training/validation loss curves
• Accuracy progression graphs
• Misclassified image analysis
• Feature correlation heatmaps
• t-SNE cluster visualization

🎓 Academic Context

Course: SIT744 Deep Learning - Deakin University
Author: Varun Kumar
Research Paper: Classification of Trash for Recyclability Status

💡 Key Techniques Demonstrated

• ✅ Transfer learning with pretrained CNNs
• ✅ Class imbalance handling through oversampling
• ✅ Strategic data augmentation
• ✅ Selective layer fine-tuning
• ✅ Early stopping to prevent overfitting
• ✅ Cross-domain generalization testing
• ✅ Feature extraction and visualization
• ✅ Performance profiling and optimization

📝 Project Highlights

1. Smart Fine-Tuning: Unfreezing only the last convolutional block achieved best results
2. Balanced Learning: Oversampling + augmentation improved accuracy by 1%
3. Real-World Testing: 81% accuracy on completely new waste images
4. Efficient Pipeline: 90% reduction in data loading time through optimization
5. Interpretability: Correlation analysis explains misclassification patterns

🔗 References

• TrashNet Dataset
• ResNet Paper
• Australian Recycling Facts

⚠️ Notes

• Designed for Google Colab with GPU support
• Requires ~2GB storage for dataset
• Training time: ~10-15 minutes on T4 GPU
• Best results with balanced + Aug2 configuration

🤝 Contributing

Contributions welcome! Feel free to:

• Add new augmentation strategies
• Test on different architectures
• Expand to more waste categories
• Improve cross-domain performance

📄 License

This project is open source and available under the MIT License.

📧 Contact

For questions or collaborations, please open an issue on GitHub.

---

Impact: This classifier can help reduce recycling contamination, saving resources and reducing environmental pollution through automated waste sorting.

Find the resources here: https://drive.google.com/drive/folders/14TOWegOtAg8Tfdjy9NRyVN0jLoLlrlA7?usp=sharing

---

Note: This project was developed as part of academic coursework. The techniques demonstrated are applicable to various NLP tasks requiring domain adaptation with limited computational resources.