用于车辆识别的深度学习模型的优化
|
摘要
在车辆识别模型中,为减小深层卷积神经网络的计算开销,对网络模型参数进行优化,基于实验确定了模型最优参数,从而以较少的网络层数获得较高的车辆识别精度.针对真实拍摄场景车辆图像尺寸较小的问题,使用复制边界的方法减小卷积过程中的像素损失,以提高识别精度.基于车辆公开数据集ImageNet和PKU-VD进行实验,并与现有的高精度模型比较,结果表明,优化后的卷积神经网络的车辆识别精度高达99.74%,优于CNN+Adaboost的97.02%和GoogLeNet-lite的99.35%.
For decreasing computing cost of deep convolutional neural network in vehicle recognition model,the parametersof the model are optimized through experiments,so as to achieve a higher vehicle recognition rate with fewer network layers.In view of the problem that the image size of the vehicle in the real shooting scene is small,the method of copying boundaryof image is used for minimizing the loss of pixels in the convolution process and improving the correct rate. Experiments arecarried out based on the ImageNet and PKU-VD vehicle public datasets,and this method is compared with existing modelswith high accuracies. The results show that the optimized three-layer convolutional neural network has a vehicle recognitionaccuracy of 99.74%,which is superior to 97.02% of CNN+Adaboost and 99.35% of GoogLeNet-lite.
For decreasing computing cost of deep convolutional neural network in vehicle recognition model,the parametersof the model are optimized through experiments,so as to achieve a higher vehicle recognition rate with fewer network layers.In view of the problem that the image size of the vehicle in the real shooting scene is small,the method of copying boundaryof image is used for minimizing the loss of pixels in the convolution process and improving the correct rate. Experiments arecarried out based on the ImageNet and PKU-VD vehicle public datasets,and this method is compared with existing modelswith high accuracies. The results show that the optimized three-layer convolutional neural network has a vehicle recognitionaccuracy of 99.74%,which is superior to 97.02% of CNN+Adaboost and 99.35% of GoogLeNet-lite.
引文
[1] COLLINS R T,LIPTON A J,KANADE T. Introduction to the specialsection on video surveillance[J]. IEEE Trans on Pattern Analysis andMachine Intelligence,2000,22(22):745-746.
[2] LONG W,YANG Y H. Stationary background generation:An alterna-tive to the difference of two images[J]. Pattern Recognition,1990,23(12):1351-1359.
[3]向川平.基于光流场的视频运动检测研究[D].成都:西华大学,2011.XIANG C P. Research on Optical Flow and Its Application on VideoMotion Detection[D]. Chengdu:Xihua University,2011(in Chinese).
[4] HU W,BAI L. A new approach for vehicle color recognition based onspecular-free image[C]//Proceedings of SPIE. The International Societyfor Optical Engineering. IEEE,2013.
[5] BEHLEY J,STEINHAGE V,CREMERS A B. Laser-based segmentclassification using a mixture of bag-of-words[C]//International Confer-ence on Intelligent Robots and Systems. IEEE,2013.
[6]徐骏骅.基于边缘检测与模式识别的车脸识别算法[J].控制工程,2018,25(2):357-361.XU J H. Car face recognition algorithm based on edge detection andpattern recognition[J]. Control Engineering of China,2018,25(2):357-361(in Chinese).
[7] HU A Q,HONG L I,ZHANG F,et al. Deep Boltzmann Machinesbased Vehicle Recognition[C]//Control and Decision Conference. IEEE,2014.
[8] SHI T,ZHANG C,LI F,et al. Application of alternating deep beliefnetwork in image classification[C]//Control and Decision Conference.IEEE,2016.
[9]康妙,计科峰,冷祥光,等.基于栈式自编码器特征融合的SAR图像车辆目标识别[J].雷达学报,2017,6(2):167-176.KANG M,JI K F,LENG X G,et al. SAR target recognition with fea-ture fusion based on stacked autoencoder[J]. Journal of Radars,2017,6(2):167-176(in Chinese).
[10] SONG X,RUI T,ZHA Z,et al. The Adaboost algorithm for vehicledetection based on CNN features[C]//International Conference on Inter-net Multimedia Computing and Service. ACM,2015.
[11] FEYZABADI S. Joint deep learning for car detection[J]. Computer Vi-sion and Pattern Recognition,2014,ar Xiv:1412.7854.
[12] PYO J,BANG J,JEONG Y. Front collision warning based on vehicledetection using CNN[C]//Soc Design Conference. IEEE,2016.
[13] SIMONYAN K,ZISSERMAN A. Very deep convolutional networks forlarge-scale image recognition[J]. Computer Vision and Pattern Recogni-tion,2014,ar Xiv:1409.1556.
[14] SZEGEDY C,LIU W,JIA Y,et al. Going deeper with convolutions[C]//IEEE Conference on Computer Vision and Pattern Recognition. IEEEComputer Society,2015.
[15] YAN K,TIAN Y,WANG Y,et al. Exploiting multi-grain ranking con-straints for precisely searching visually-similar vehicles[C]//IEEE Inter-national Conference on Computer Vision. IEEE Computer Society,2017.
[2] LONG W,YANG Y H. Stationary background generation:An alterna-tive to the difference of two images[J]. Pattern Recognition,1990,23(12):1351-1359.
[3]向川平.基于光流场的视频运动检测研究[D].成都:西华大学,2011.XIANG C P. Research on Optical Flow and Its Application on VideoMotion Detection[D]. Chengdu:Xihua University,2011(in Chinese).
[4] HU W,BAI L. A new approach for vehicle color recognition based onspecular-free image[C]//Proceedings of SPIE. The International Societyfor Optical Engineering. IEEE,2013.
[5] BEHLEY J,STEINHAGE V,CREMERS A B. Laser-based segmentclassification using a mixture of bag-of-words[C]//International Confer-ence on Intelligent Robots and Systems. IEEE,2013.
[6]徐骏骅.基于边缘检测与模式识别的车脸识别算法[J].控制工程,2018,25(2):357-361.XU J H. Car face recognition algorithm based on edge detection andpattern recognition[J]. Control Engineering of China,2018,25(2):357-361(in Chinese).
[7] HU A Q,HONG L I,ZHANG F,et al. Deep Boltzmann Machinesbased Vehicle Recognition[C]//Control and Decision Conference. IEEE,2014.
[8] SHI T,ZHANG C,LI F,et al. Application of alternating deep beliefnetwork in image classification[C]//Control and Decision Conference.IEEE,2016.
[9]康妙,计科峰,冷祥光,等.基于栈式自编码器特征融合的SAR图像车辆目标识别[J].雷达学报,2017,6(2):167-176.KANG M,JI K F,LENG X G,et al. SAR target recognition with fea-ture fusion based on stacked autoencoder[J]. Journal of Radars,2017,6(2):167-176(in Chinese).
[10] SONG X,RUI T,ZHA Z,et al. The Adaboost algorithm for vehicledetection based on CNN features[C]//International Conference on Inter-net Multimedia Computing and Service. ACM,2015.
[11] FEYZABADI S. Joint deep learning for car detection[J]. Computer Vi-sion and Pattern Recognition,2014,ar Xiv:1412.7854.
[12] PYO J,BANG J,JEONG Y. Front collision warning based on vehicledetection using CNN[C]//Soc Design Conference. IEEE,2016.
[13] SIMONYAN K,ZISSERMAN A. Very deep convolutional networks forlarge-scale image recognition[J]. Computer Vision and Pattern Recogni-tion,2014,ar Xiv:1409.1556.
[14] SZEGEDY C,LIU W,JIA Y,et al. Going deeper with convolutions[C]//IEEE Conference on Computer Vision and Pattern Recognition. IEEEComputer Society,2015.
[15] YAN K,TIAN Y,WANG Y,et al. Exploiting multi-grain ranking con-straints for precisely searching visually-similar vehicles[C]//IEEE Inter-national Conference on Computer Vision. IEEE Computer Society,2017.