Adaptive Fine-tuning for Deep Transfer Learning Based Traffic Signs Classification

In recent years, traffic signs recognition represents an important issue in intelligent transportation systems. Several systems use traffic signs recognition including, driving assistance systems, road safety, autonomous driving. The traffic signs recognition aims to read and interpret road signs to inform the driver if he could not see them or when the vehicle is in self-driving mode. Each category of traffic sign has a special shape and color. This includes regulatory, warning, and guide signs. This paper proposes a practical solution for traffic signs classification based on convolutional neural networks technique to classify input images into 43 classes. Also, this paper provides a comparison between the Support Vector Machine (SVM) and the Softmax classifier. We have analyzed the impact of fine-tuning the pre-trained CNN model in the transfer learning algorithm. As a result, the SVM classifier in CNN achieves an accuracy of 96.60%, whereas the Softmax classifier accuracy is 97.84%. Experimental results demonstrate that fine-tuning in transfer learning can lead to significant performances in terms of accuracy of classification.