Research on Methods of Vehicle Image Retrieval in Expressway Scenarios Based on Deep Learning Neural Network

The vehicle image retrieval in highway scenes is influenced by the complexity and diversity of vehicle types, as well as the impact of lighting and climate changes on vehicle image clarity. The paper aims to achieve high-precision vehicle retrieval in complex scenes on highways. Firstly, the paper constructs a vehicle image retrieval dataset of Southeast University (SEU) including whole vehicle image test set, vehicle face image test set, vehicle window test set, vehicle license plate test set. Secondly, three variants of CNN models are investigated for the vehicle image retrieval in expressway scenarios, including VGG16-VIR-ES, ResNet50-VIR-ES and DenseNet121-VIR-ES. Finally, based on vehicle image retrieval dataset of Southeast University (SEU), the research of compare experiment on VGG16-VIR-ES, ResNet50-VIR-ES and DenseNet121-VIR-ES are carried out. The experimental results show that DenseNet121-VIR-ES outperform the other models, and the accuracy of the vehicle image retrieval method based on DenseNet121 network is 92.12%, 91.55%, 92.86% and 84.98% for the whole vehicle image test set, vehicle image test set, window image test set and license plate image test set, respectively.