This project is part of a Computer Vision course.
In this project, I built several neural networks in order to classify the SVHN dataset that contains images of house numbers (0-9).
I compare between a Fully Connected Network and a Convolutional Neural Network for the task of image classification (spoiler: CNN is much better).
I also show how regularization and data data augmentation improve the results of the network.
SVHN is a real-world image dataset for developing machine learning and object recognition algorithms with minimal requirement on data preprocessing and formatting.
- 10 classes, 1 for each digit. Digit '1' has label 1, '9' has label 9 and '0' has label 0.
- 73257 digits for training, 26032 digits for testing, and 531131 additional, somewhat less difficult samples, to use as extra training data