RealWaste Image Classification (PyTorch)

A deep learning project for classifying waste images using a Convolutional Neural Network (CNN) in PyTorch. This project uses the RealWaste dataset from the UCI Machine Learning Repository and demonstrates data preprocessing, augmentation, model training, evaluation, and visualization.

Figure: Sample image with its label for all 9 classes.

Project Structure

RealWaste CNN/
├── main.ipynb           # Jupyter notebook with all code
├── README.md            # Project documentation
├── checkpoint.pth       # Model checkpoint
├── history.json         # Training history
├── dataset/             # Original dataset
├── dataset_split/       # Train/val/test splits

Dataset

• Source: UCI RealWaste Dataset
• Classes: Cardboard, Food Organics, Glass, Metal, Miscellaneous Trash, Paper, Plastic, Textile Trash, Vegetation
• Preprocessing: Images are split into train/val/test folders and augmented to balance class counts.

Setup

1. Clone the repository:

   git clone https://github.com/InduwaraGunasena/RealWaste-Image-Classification.git
   cd RealWaste-Image-Classification

2. Install dependencies:
   • Python 3.8+
   • PyTorch
   • torchvision
   • scikit-learn
   • matplotlib
   • tqdm
   • pandas
   • seaborn
   • Jupyter Notebook

   pip install torch torchvision scikit-learn matplotlib tqdm pandas seaborn jupyter

3. Download the RealWaste dataset and place it in the dataset/realwaste-main/RealWaste directory.

Usage

Open main.ipynb in Jupyter Notebook and run the cells sequentially:

1. Environment Setup:
   • Detects GPU and sets device.
2. Data Preparation:
   • Splits dataset into train/val/test.
   • Augments training data for class balance.
   • Optionally caches datasets for faster loading.
3. Data Visualization:
   • Shows sample images and class distributions.
4. Model Definition:
   • Defines a custom CNN architecture with batch normalization, dropout, and global average pooling.
5. Training:
   • Trains the model with progress bars showing loss and accuracy.
   • Supports early stopping and checkpointing.
6. Evaluation:
   • Plots training history (loss, accuracy, precision, recall).
   • Evaluates on test set and prints metrics.
   • Displays confusion matrix and classification report.
7. Prediction Visualization:
   • Shows grid of test images with predicted/true labels and probability distributions.

Results & Visualization

• Training History: Plots for loss, accuracy, precision, and recall over epochs.
• Test Predictions: Visual grids showing model predictions and confidence for test images.
• Confusion Matrix: Visualizes class-wise performance.
• Classification Report: Detailed metrics for each class.

Model Architecture & Performance

Our custom CNN model is designed for robust image classification with the following layer-wise structure:

• Input Layer: Accepts RGB images of size 128x128.
• Block 1: Two convolutional layers (64 filters), each followed by batch normalization and ReLU activation, then max pooling and dropout.
• Block 2: Two convolutional layers (128 filters), batch normalization, ReLU, max pooling, and dropout.
• Block 3: Two convolutional layers (256 filters), batch normalization, ReLU, max pooling, and dropout.
• Block 4: One convolutional layer (256 filters), batch normalization, ReLU, and dropout.
• Global Average Pooling: Reduces spatial dimensions to a single value per channel.
• Fully Connected Layers:
  • Dense layer (128 units) with ReLU and dropout
  • Output layer (for 9 classes)

This architecture uses batch normalization and dropout throughout to improve generalization and training stability.

Model Summary:

The model contains multiple convolutional blocks, global average pooling, and fully connected layers, resulting in a compact and efficient design for multi-class waste classification.

Figure: Layer-wise summary of the CNN architecture.

After 50 epochs of training, the model achieved:

• Accuracy: 80.47%
• Precision: 80.12%
• Recall: 81.47%

Training History:

The training and validation curves show steady improvement and convergence, with minimal overfitting due to regularization.

Figure: Training and validation metrics over epochs.

Confusion Matrix:

The confusion matrix below visualizes the model's performance across all classes, highlighting areas of strong and weak classification.

Figure: Confusion matrix for test set predictions.

Customization

• Change batch_size, img_size, or model architecture in main.ipynb as needed.
• Adjust augmentation pipeline for different image sizes or transformations.

Citation

If you use this project or dataset, please cite the UCI RealWaste dataset and this repository.