This Jupyter Notebook explores emotion detection using various Convolutional Neural Network (CNN) techniques, focusing on improving accuracy through custom CNNs, data augmentation, and transfer learning.
- Custom CNN: The initial model was built using normalization, max-pooling, and ReLU activation functions, achieving 56% accuracy due to imbalanced data.
- Data Augmentation: To address the data imbalance, I applied techniques like rotation, flip, and zoom, leading to more diverse training data.
- Transfer Learning with VGG16: Implemented transfer learning by fine-tuning a VGG16-based model, resulting in significant performance improvements.
- ResNet for Optimal Accuracy: Ultimately, I used the ResNet architecture, achieving a remarkable 92% accuracy for emotion detection.
- Code for building and training a custom CNN, with detailed explanations of layers and functions used.
- Implementation of data augmentation strategies to boost model robustness.
- Steps to fine-tune a VGG16, ResNet model and evaluate its performance.
- Instructions for using the ResNet model to achieve high accuracy.
- Clone or download the repository.
- Install the necessary dependencies.
- Follow the step-by-step instructions in the Jupyter Notebook to build, train, and evaluate the models.
- Observe the differences in accuracy across the custom CNN, VGG16, 5C7F and ResNet models.
- Python 3.10
- Jupyter Notebook
- Deep learning libraries like TensorFlow, Keras
- Data visualization libraries such as Matplotlib, Computer Vision