Traffic sign classification network using inception module

Abstract

With the rapid development of the automobile industry, the demand for autonomous driving becomes more and more urgent, and the traffic sign recognition technology in autonomous driving is an indispensable technology. This paper proposes a GoogLeNet based convolutional neural network for traffic signs. This convolutional neural network improves each of the underlying Inception Modules and adds the Batch Normalization layer, effectively avoiding over-fitting of the network. We use a sparse structure that conforms to the Hebbain principle to reduce the parameters and improve the generalization ability of the network, which can extract the features of the image more accurately. Meanwhile, the network also reduces the parameters of the full connection layer by 20 times through the continuous two-layer pooling layer, which greatly speeds up the network training. Finally, the network is trained using the GTSRB data set and the classification accuracy rate can reach 98%. At the same time, we also verified the validity of the network on the MNIST dataset and the pneumonia dataset. The classification accuracy rate can reach 100% on the above two datasets. Experimental results on the above data sets show the validity of the convolutional neural network.