Waste Classification Using Convolutional Neural Networks (CNN)

This project focuses on developing a simple convolutional neural network (CNN) to classify waste using the RealWaste dataset. The dataset includes nine classes of waste to train the model, enabling applications like automated waste sorting and environmental monitoring.

Dataset

We utilized the RealWaste dataset, which includes the following classes:

• Cardboard
• Food Organics
• Glass
• Metal
• Miscellaneous Trash
• Paper
• Plastic
• Textile Trash
• Vegetation

The dataset was split into:

• Training set: 60%
• Validation set: 20%
• Testing set: 20%

Dataset link: RealWaste Dataset

Project Features

1. Custom CNN Model

We implemented a custom CNN architecture with the following specifications:

• Convolutional Layers: Four layers with ReLU activation and varying filter sizes (64, 128, 256, and 512).
• Pooling Layers: Max pooling for down-sampling.
• Fully Connected Layers:
  • First layer: 512 neurons with ReLU and 50% dropout.
  • Output layer: 9 neurons with softmax activation.
• Optimizer: Adam (chosen over SGD for its faster convergence).
• Loss Function: Sparse categorical cross-entropy.

2. Fine-tuning Pretrained Model

We fine-tuned the ResNet50 model using transfer learning. This pretrained model significantly outperformed the custom model in terms of accuracy, precision, and recall.

3. Hyperparameter Tuning

We experimented with different learning rates (0.1, 0.01, 0.001, and 0.0001). The best performance was achieved with a learning rate of 0.0001.

Evaluation

Custom Model Performance

• Test Accuracy: 0.1987
• Precision: 0.0395
• Recall: 0.1987

Fine-tuned ResNet50 Performance

• Test Accuracy: 0.8875
• Precision: 0.8913
• Recall: 0.8875

Advantages and Limitations

Custom Model

• Advantages:
  • Tailored to the dataset.
  • Lightweight and flexible.
• Limitations:
  • Requires more data for effective training.
  • Longer training time.
  • Limited generalization.

Pretrained Model (ResNet50)

• Advantages:
  • Leverages knowledge from large-scale datasets.
  • Performs well with limited data.
  • Faster convergence.
• Limitations:
  • Higher computational resource requirements.
  • Potential overfitting if not fine-tuned carefully.

Results

Fine-tuning the ResNet50 model resulted in significantly better performance than the custom CNN model, highlighting the advantages of transfer learning in image classification tasks.