Athens Institute for Education and Research ATINER ATINER ' s Conference Paper Series TRA 2017 - 2275

Vehicle Manufacturer Recognition (VMR) has recently become a key feature to be used when license plate numbers cannot be detected or fake plate numbers are used in traffic control and management applications. One of the clearest indicators of a vehicle manufacturer is its vehicle logo. In this way, most of the vision-based VMR approaches are based on vehicle logo classification. In this work, we contribute to this topic by presenting an experimental comparison of different Convolutional Neural Networks (CNN) applied to vehicle logo recognition. A data set containing 26.289 logo images, corresponding to 30 car manufacturers and one non-logo class, is used to validate the proposed approach. The ability of the CNN to learn features robust to rotations is analyzed by artificially increasing the data set up to 446.913 adding rotated samples. The proposed methodology is also compared with a baseline classifier based on Histograms of Oriented Gradients (HOG) and Support Vector Machines (SVM). Experiments show an average accuracy of 99.87% in the validation set, which clearly demonstrate the potential of our CNN-based logo classifier to deal with VMR.