Effect of Classifiers in Traffic Signs Classification using Convolutional Neural Network

Abstract

We consider the problem of traffic sign classification through two different neural network architectures. In particular, images of traffic signs with different quality are utilized as an input to a specific neural network and the network predict their corresponding classes. There are different architectures of neural networks and in general, they perform well for different problems. One such group of architectures is the convolution neural network (CNN) that is considered as an efficient tool for image classification. In this thesis, we used a CNN classifier called AlexNet that has a specific architecture and is well used for image classification. The AlexNet predicts a label for the object in the image along with the probabilities of other object labels. Since it is possible to replace the final layers of Alexnet with some other classifier, we consider the use of support vector machine (SVM) for image classification together with the features identified from AlexNet. The performance of the two frameworks, AlexNet with built-in SoftMax classifier and AlexNet with SVM, are compared for different values of dropouts in the fully connected layers of AlexNet. German Traffic Signs Recognition Benchmark (GTSRB) is used for training and validation of the network and it is observed that we can reach to the accuracy of 94.78% with SoftMax and 97% with SVM classifier for the selected data and classes.