Deep Pre-trained Models for Computer Vision Applications: Traffic sign recognition

Abstract

Objects detection and Recognition are an important task for computer vision field and intelligent transportation systems. Generally, these tasks remain challenging for the artificial machines due to the need of pre-learning phase in which the machine acquires an intelligent brain. Some researchers have shown that deep learning tools work well in computer vision, image processing, and pattern recognition. To solve such tasks, this paper focuses on deep Convolutional Neural Network (CNN) and its architectures, such as, VGG16, VGG19, AlexNet, and Resnet50. An overview for the techniques and schemes used for computer vision applications such as Road Sign Recognition will be introduced. Then by customizing the hyperparameters for each pre-trained models, we re-implement these models for the traffic sign recognition application. In the experiments, these pre-trained CNN classifiers are trained and tested with the German Traffic Sign Recognition Benchmark dataset (GTSRB). Experimental results show that the proposed scheme achieved a good performance results in terms of evaluations metrics of traffic signs recognition. A performance comparison analysis between the selected pre-trained models for traffic sign recognition confirmed that the AlexNet model outperforms all other implemented models.