Instead of using Feed-Forward Neural Networks as in the previous assignment, this exercise aims to implement a Convolutional Neural Network (CNN), for solving two image classification problems: German Traffic Sign Recognition Benchmark and Cifar-100.
The base implementation in Python with the tensorflow library is given and the aim is to decide on the following factors:
- Architecture of your net.
- Optimization parameters and algorithms to train the net.
- Performance improvement (regularization, data augmentation, etc.)
For the German Traffic Sign Recognition Benchmark, multiple classic CNN architectures were tested, including LeNet, AlexNet and VGG-16 pre-trained. On the other side, a custom architecture was used for classifying the Cifar-100 dataset, presented below.
The best performance found for the German Traffic Sign Recognition Benchmark was using the model AlexNet without any pretraining, with little difference compared to the VGG-16. The validation accuracy was of 95.56%.
The custom architecture developed for the Cifar-100 achieved a 50.08% of validation accuracy, much better than in previous assignment with 27.88%.
- Student Name 1: Stefano Baggetto
- Student Name 2: Giorgio Segalla
- Student Name 3: Angel Igareta (angel@igareta.com)