Abstract: Fine-grained vehicle identification is challenging, especially when the two vehicles differ in appearance and subtleness. However, the general neural network model ignores the connection between the additional attributes. This paper proposes a convolution neural network based on improved triplet loss training for vehicle multi-attribute learning, which is used to implement fine-grained vehicle identification. Specifically, by changing the structure of the traditional neural network, the vehicle identification problem is transformed into a multi-attribute learning problem. In this paper, the triplet loss function is improved to train the network to achieve fine-grained vehicle identification. At the same time, it creates a multi-attribute vehicle data set and completes the training work. The results show the potential of the method.

Key words: fine-grained vehicle recognition, vehicle multi-attribute, multi-task learning, convolution neural network, metric learning, vehicle multi-attribute data

摘要： 细粒度车辆识别极具挑战性，尤其在两辆车的外型差异及其细微的时候。通过车辆的附加属性能够提高车辆识别效果，但一般的神经网络模型忽略了附加属性间的联系，提出一种基于改进的triplet loss作为损失函数的车辆多属性学习的卷积神经网络，用于实现细粒度车辆多属性识别。具体而言，通过对传统神经网络结构的改变，将车辆识别问题转化为多属性学习问题。对三元组损失函数进行改进用于训练网络以实现细粒度车辆识别。同时，创建了一个车辆多属性数据集并完成训练工作，结果显示了该方法的潜力。

关键词: 细粒度车辆识别, 车辆多属性, 多任务学习, 卷积神经网络, 度量学习, 车辆多属性数据集