Traffic Sign Recognition by Integrating Convolutional Neural Network and Support Vector Machine

This paper represents a combined model of convolutional neural network (CNN) and support vector machine (SVM) for traffic sign recognition. This model was built by training a CNN model. Once the CNN model is fully trained the output from the later layers of CNN can be used as features. These features were then fed into SVM for classification purpose. Three different models of CNN: modified version of LeNet, AlexNet and ResNet-50 were considered to build three CNN-SVM models. The integrated model of Resnet50 and SVM seems to perform better than ResNet-50 while the other two merged models of Lenet and Alexnet performed worse than their corresponding CNN models. One reason of this can be ResNet-50 having a shallow classification part consisting of only one fully connected layer while modified version of LeNet and AlexNet have 3 and 4 fully connected layers respectively. This combined approach provides for a good comparison between SVM and CNN as classifiers since the features used in both these classifiers are same. So a comparative analysis among three different CNN models and their corresponding integrated models is shown. In our analysis, we considered different measurement metrices like accuracy, precision, recall and F1 score. We used German Traffic Sign Detection Benchmark (GTSRB) dataset. This dataset gives access to a wide range of traffic sign images.