A Deep Learning Network for Vehicle Identification: Dense-Inception

For future intelligent transportation, vehicle classification can provide driver assistance systems, traffic flow statistics, and intelligent parking systems. With the increase in the number of vehicles in cities, it is necessary to develop effective vehicle classification methods for better traffic control, intelligent parking, and electronic toll collection. In order to further optimize the accuracy of identification, this study is based on a self-collected dataset. Based on existing networks, the study adopts the Inception module and the DenseNet network structure. The Inception module combines different convolutional layers in parallel before the image enters the network. Before entering the network, an STN layer is added to preliminarily filter out interference in the image using the attention mechanism. Finally, before entering the fully connected layer, an SPP layer is added to construct a new network structure: Dense-Inception. Due to the relatively small overall amount of data collected, in order to increase the training volume, this project uses mixup and color augmentation, and color inversion to process the original images, further enriching our own dataset. To verify the practical effect of this research, this project compared various deep learning networks that performed well and found that compared to those networks, the network in this research showed significant improvement.