基于深度迁移学习的车辆信息识别方法  

Vehicle information recognition method based on deep migration learning

作　　者：童强 李太君[1,2] 刘笑嶂[1] TONG Qiang;LI Taijun;LIU Xiaozhang(School of information science and technology, Hainan University, Haikou 570228, China;Provincial Key Internet Information Retrieval Laboratory, Hainan University, Haikou 570228, China)

机构地区：[1]海南大学信息科学技术学院,海南海口570228 [2]海南大学省重点Internet信息检索实验室,海南海口570228

出　　处：《电视技术》2019年第1期66-71,共6页Video Engineering

基　　金：国家自然科学基金项目(61562017)

摘　　要：针对在深度卷积神经网络中存在对样本数据需求量大和过拟合的问题,提出一种基于深度网络迁移学习的车辆信息识别方法。该方法通过在ImageNet数据集上预训练的深度网络VGG-19进行同构空间下的特征迁移;结合改进的模型损失函数Softmax搭建网络全连接层,并冻结中低层卷积层、利用不同学习率来微调高层卷积层和全连接层参数。然后利用车辆数据集进行实验验证,结果表明该方法能在训练精度与测试精度上有较高的准确识别率,其中测试准确识别率达到97.73%;同时解决了样本数据不足带来的模型过拟合的问题,具有良好的鲁棒性。Aiming at the problem of large and over-fitting of sample data in deep convolutional neural networks,a vehicle information recognition method based on deep network migration learning is proposed.In this method,the feature migration under isomorphic space was carried out by the deep network vgg-19 pre-trained on the ImageNet data set.and combines the improved model loss function Softmax to build the network all connected layer,and freeze the middle and low layer coiling layer,and use different learning rates to adjust the high level convolution layer and the full connection layer parameters.Then the vehicle data set is used to verify the experimental results.The results show that the method has a higher accurate recognition rate in training accuracy and test accuracy,and the accurate recognition rate of the test reaches 97.73%.At the same time,the problem of model over-fitting caused by insufficient samples data is solved,which has good robustness.

关 键 词：深度学习 迁移学习 损失函数 车辆信息识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]