Abstract: The fine-grained vehicle model is one of the main clues for vehicle recognition and an important part of the intelligent transportation system. A multi-scale feature fusion Convolutional Neural Network（CNN） model is proposed, which aims to solve the recognition difficulty caused by the wide variety of vehicle models and the complex and changeable environment. This method is based on the traditional CNN. It can accomplish the fine-grained recognition of the vehicle model by extracting and fusing vehicle features from the bottom layer and the top layer of the network. Compared with other recognition methods based on CNN, the proposed method can significantly reduce the size of network parameters while improving the classification accuracy. Using the public dataset CompCars, the experimental results show that the method reaches a recognition accuracy of 98.43%, and the model parameter size is only 3.93 MB. The average recognition time for each image is 0.83 ms.

Key words: fine-grained recognition of vehicle model, Convolutional Neural Network（CNN）, multi-scale feature fusion, deep learning

摘要： 车辆精细型号是车辆识别的主要线索之一，也是智能交通系统的重要组成部分。针对车辆精细型号种类繁多、车辆所处环境复杂多变等因素，提出一种基于多尺度特征融合的车辆精细型号识别方法。该方法基于传统的卷积神经网络，通过提取并融合来自网络底层和高层的车辆特征，完成对车辆精细型号的识别。与其他基于卷积神经网络的车辆精细型号识别方法相比，该方法在提高分类准确率的同时还大幅度降低了整体网络的参数规模。实验结果表明，在公开数据集CompCars的监控场景下其识别精度达到了98.43%，且模型参数大小仅为3.93 MB，平均每张图片只需0.83 ms的分类时间。

关键词: 车辆精细型号识别, 卷积神经网络, 多尺度特征融合, 深度学习