Accuracy Comparison of CNN Networks on GTSRB Dataset

Abstract

In this era, interpreting and processing the data of traffic signs has crucial importance for improving autonomous car technology. In this respect, the relationship between the recognition of traffic signs and industrial applications is highly relevant. Although real-world systems have reached that related market and several academic studies on this topic have been published, regular objective comparisons of different algorithmic approaches are missing due to the lack of freely available benchmark datasets. From this point of view, we compare the AlexNET, DarkNET-53, and EfficientNET-b0 convolutional neural network (CNN) algorithms according to validation performance on the German Traffic Signs Recognition Benchmark (GTSRB) dataset. Considering the equal training and test conditions 70% of data as training, 15% of data as training validation, and 15% of data were chosen as test data. Experimental results show us that EfficientNETb0 architecture has 98.64%, AlexNET architecture has 97.45% and DarkNet-53 architecture has 94.69% accuracy performance.