Multi-View Image-based Vehicle Brand Recognition System with Cascaded Convolutional Neural Network

초록

In this work, we propose a novel system to classify vehicle brand using the cascaded convolutional neural network (CNN). While conventional methods use only the front image of a vehicle, we can use input images in all directions and recognize the characteristic. In the simulation results, we achieve high performance of 91% accuracy using the cascaded CNN. We also demonstrate the distinction of our system compared to conventional network using classification activation map for visualizing the region that the network focuses on.