Training a Multi-Scale Convolutional Neural Network to perform multi-class classification on the German Traffic Sign Recognition Benchmark. The dataset contains RGB images from 43 classes of traffic signs with different resolutions. The Architecture of the classifier takes inspiration from the 2011 paper by Pierre Sermanet and Yann LeCun. It uses Data Augmentation (shear, rotation, translation, zoom, elastic transform, gaussian noise) to increase model accuracy.
The Best Accuracy achieved is 98.75% on the test set (classifier.h5) and is similar to human perfomance (recorded at 98.8% for this dataset).
