Data Augmentation and Deep Learning Applied for Traffic Signs Image Classification

Abstract

Traffic sign recognition is an essential phase for intelligent autonomous driving systems. In this work, we have presented a recognition solution which is based on the classification performed by supervised learning method through deep learning an artificial using Convolutional Neural Network (CNN). Three pre-trained neural networks such as, GoogLeNet, VGG-16, AlexNet, were adapted for transfer learning algorithm. In order to improve the classification accuracy and to solve the problem of the insufficient amount of training dataset or the unequal number of images for each class, we have used the data augmentation technique applied on German Traffic Sign Benchmark dataset (GTSRB). The obtained experimental results show that the size of the dataset as the number of training images has a significant impact on the classification accuracy of the CNN model. The number of images was increased from 51,840 to 96,750 images by using data augmentation technique. Therefore the classification accuracy was increased from 86.72% to 97.33%.