基于深度卷积神经网络的车型识别方法

针对现有车辆车型视觉识别技术中的检测精度不高、难以适应天气环境变化、难以从视频图像中准确提取出用于识别的车辆图像、难以对车辆车型子类进行识别分类、难以兼顾识别精度和检测效率等不足，将深度卷积神经网络引入车辆目标定位、识别和分类（子类）问题中.利用深度卷积神经网络自动完成车型的深度特征学习，在特征图上进行逻辑回归，从道路复杂背景中提取出感兴趣区域；利用softmax分类器训练特征实现车型识别；为了优化softmax在深度卷积神经网络分类过程中出现的类内间距大的问题，引入中心损失函数对softmax损失函数进行优化，提高类间分散性与类内紧密性.在BIT-Vehicle车型数据集中的实验结果显示，提出方法的平均精度为89.67%，检测和识别时间为159 ms；与传统的分类方法相比，识别精度提高约20%，效率提高10倍以上，检测鲁棒性有明显提升；与未改进前的深度卷积神经网络相比，检测精度提高0.6%，速度提高0.29倍.

Vehicle category recognition based on deep convolutional neural network

A novel detection model was proposed to solve the problems of low detection accuracy, difficulty in adapting to changeable climate environment, extracting the vehicles from the video accurately, meeting fine classification based on visual method, and giving attention to both the recognition accuracy and the detection efficiency and so on in the existing vehicle identification and classification. The problems contain three consecutive stages:vehicle detection, features extraction and classification based on deep convolutional neural network. The deep learning network was used to automatically complete vehicle feature extraction and logical regression was performed on the feature map to extract the region of interest from the complex background. Then vehicle recognition was implemented to train the existing features through a softmax classifier. The central loss function was introduced to optimize this problem and improve inter-class dispension and intra-class compactness as much as possible in order to optimize the problem that softmax will cause large class spacing in deep convolutional neural network. Experiments on BIT-Vehicle dataset demonstrated that the identification average accuracy of our algorithm was about 89.67%, and the recognition rate was about 159 ms. The recognition accuracy increased by approximately 20% compared with conventional machine learning methods, and the rate increased by at least 10 times. Robustness was significantly improved. The recognition accuracy increased by approximately 0.6% and the rate increased by at least 0.29 times compared with the unimproved deep convolution neural network.