复杂背景下车型识别分类器
|
摘要
细粒度车型图像的类间特征差异小,在复杂图片背景中识别干扰因素多。为提高模型在复杂背景中对图像的特征提取能力和识别准确度,提出了基于支持向量机(SVM)和深度卷积神经网络(DCNN)的分类器集成模型Softmax-SVM。它将交叉熵代价函数与hinge损失函数相结合,代替Softmax函数层,减少了过拟合的发生。同时,设计了一个10层的DCNN提取特征,避免了手工提取特征的难题。实验数据集为复杂背景下的27类精细车型图像,尤其还包含同一汽车厂商的相近车型。实验结果表明,在不进行大量预处理的前提下,Softmax-SVM分类器识别269张测试样本能够获得97.78%的准确率,识别每张样本的时间为0.759s,明显优于传统模式识别方法和未改进前的DCNN模型。因此,基于DCNN的Softmax-SVM分类器能够适应环境的复杂变化,兼顾识别精度与效率,为复杂背景下的细粒度车型分类提供了实际参考价值。
The feature difference among the images of fine-grained vehicle models is small and there exist many factors disturbing recognition under complex image background.To improve the feature extraction ability and the recognition accuracy of images under complex background,a classifier named Softmax-SVM is proposed based on deep convolutional neural network(DCNN)and support vector machine(SVM),in which the cross-entropy cost function is combined with the hinge loss function to replace the Softmax function layer,so that the over-fitting is avoided.Meanwhile,a 10-layer DCNN is designed to extract features automatically and the problem of manual extraction of features is also avoided.The experimental dataset consists of the images of 27 types of fine-gained vehicle models under complex background,especially of the similar models from the same car manufacturer.The experimental results show that the Softmax-SVM classifier can be used to recognize the 269 sample images without much emphasis on the pre-processing stages,and in the identification process,the accuracy rate is 97.78% and the time to identity each image is 0.759 s.The above model performs more efficiently than the traditional recognition methods and the unimproved DCNN models.In consequence,the Softmax-SVM classifier based on DCNN can adapt to the complex changes of environment and give consideration to both the recognition accuracy and efficiency,which provides practical reference value in the classification field of fine-gained vehicle models under complex background.
The feature difference among the images of fine-grained vehicle models is small and there exist many factors disturbing recognition under complex image background.To improve the feature extraction ability and the recognition accuracy of images under complex background,a classifier named Softmax-SVM is proposed based on deep convolutional neural network(DCNN)and support vector machine(SVM),in which the cross-entropy cost function is combined with the hinge loss function to replace the Softmax function layer,so that the over-fitting is avoided.Meanwhile,a 10-layer DCNN is designed to extract features automatically and the problem of manual extraction of features is also avoided.The experimental dataset consists of the images of 27 types of fine-gained vehicle models under complex background,especially of the similar models from the same car manufacturer.The experimental results show that the Softmax-SVM classifier can be used to recognize the 269 sample images without much emphasis on the pre-processing stages,and in the identification process,the accuracy rate is 97.78% and the time to identity each image is 0.759 s.The above model performs more efficiently than the traditional recognition methods and the unimproved DCNN models.In consequence,the Softmax-SVM classifier based on DCNN can adapt to the complex changes of environment and give consideration to both the recognition accuracy and efficiency,which provides practical reference value in the classification field of fine-gained vehicle models under complex background.
引文
[1]Zhang J,Zhang T,Yang Z L,et al.Vehicle model recognition method based on deep convolutional neural network[J].Transducer and Microsystem Technologies,2016,35(11):19-22.张军,张婷,杨正瓴,等.深度卷积神经网络的汽车车型识别方法[J].传感器与微系统,2016,35(11):19-22.
[2]Zheng H L,Fu J L,Mei T,et al.Learning multiattention convolutional neural network for finegrained image recognition[C]∥2017 IEEEInternational Conference on Computer Vision(ICCV),October 22-29,Venice,Italy.New York:IEEE,2017:5219-5227.
[3]Yan Y,Ni B,Yang X.Fine-grained recognition via attribute-guided attentive feature aggregation[C]∥The 25th ACM International Conference on Multimedia,October 23-27,2017,California,USA.New York:ACM,2017:1032-1040.
[4]Huang K,Zhang B L.Fine-grained vehicle recognition by deep convolutional neural network[C]∥2016 9th International Congress on Image and Signal Processing,Biomedical Engineering and Informatics,October 15-17,Datong,China.New York:IEEE,2016:465-470.
[5]Khusnuliawati H,Fatichah C,Soelaiman R.Multifeature fusion using SIFT and LEBP for finger vein recognition[J].Telecommunication Computing Electronics and Control,2017,15(1):478-485.
[6]Murtaza F,Yousaf M H,Velastin S A.Multi-view human action recognition using 2D motion templates based on MHIs and their HOG description[J].IETComputer Vision,2016,10(7):758-767.
[7]Ling Y G,Hu W P.Vehicle type recognition based on SURF and integral channel features[J].Video Engineering,2016,40(7):139-143.凌永国,胡维平.基于SURF特征与积分通道特征的车型识别[J].电视技术,2016,40(7):139-143.
[8]Fang J,Zhou Y,Yu Y,et al.Fine-grained vehicle model recognition using a coarse-to-fine convolutional neural network architecture[J].IEEE Transactions on Intelligent Transportation Systems,2017,18(7):1782-1792.
[9]Bengio Y,Courville A,Vincent P.Representation learning:A review and new perspectives[J].IEEETransactions on Pattern Analysis and Machine Intelligence,2013,35(8):1798-1828.
[10]Zou Y B,Zhou W L,Chen X Z.Research of laser vision seam detection and tracking system based on depth hierarchical feature[J].Chinese Journal of Lasers,2017,44(4):0402009.邹焱飚,周卫林,陈向志.基于深度分层特征的激光视觉焊缝检测与跟踪系统研究[J].中国激光,2017,44(4):0402009.
[11]Zhou Y C,Xu T Y,Zheng W,et al.Classification and recognition approaches of tomato main organs based on DCNN[J].Transactions of the Chinese Society of Agricultural Engineering,2017,33(15):219-226.周云成,许童羽,郑伟,等.基于深度卷积神经网络的番茄主要器官分类识别方法[J].农业工程学报,2017,33(15):219-226.
[12]Song J,Kim H I,Yong M.Fast and robust face detection based on CNN in wild environment[J].Journal of Korea Multimedia Society,2016,19(8):1310-1319.
[13]Liu B,Yu X C,Zhang P Q,et al.A semi-supervised convolutional neural network for hyperspectral image classification[J].Remote Sensing Letters,2017,8(9):839-848.
[14]Qu L,Wang K R,Chen L L,et al.Fast road detection based on RGBD images and convolutional neural network[J].Acta Optica Sinica,2017,37(10):1010003.曲磊,王康如,陈利利,等.基于RGBD图像和卷积神经网络的快速道路检测[J].光学学报,2017,37(10):1010003.
[15]Cortes C,Vapnik V.Support-vector networks[J].Machine Learning,1995,20(3):273-297.
[16]Elleuch M,Maalej R,Kherallah M.A new design based-SVM of the CNN classifier architecture with dropout for offline Arabic handwritten recognition[J].Procedia Computer Science,2016,80:1712-1723.
[17]Li X Q,Zhang Y,Liao D.Mining key skeleton poses with latent SVM for action recognition[J].Applied Computational Intelligence and Soft Computing,2017,2017:1-11.
[18]Cheng L Y,Mi G Y,Li S,et al.Quality diagnosis of joints in laser brazing based on principal component analysis-support vector machine model[J].Chinese Journal of Lasers,2017,44(3):0302004.程力勇,米高阳,黎硕,等.基于主成分分析-支持向量机模型的激光钎焊接头质量诊断[J].中国激光,2017,44(3):0302004.
[19]Huang F J,LeCun Y.Large-scale learning with SVMand convolutional for generic object categorization[C]∥2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition(CVPR′06),June 17-22,New York,USA.New York:IEEE,2006:284-291.
[20]Lapin M,Hein M,Schiele B.Loss functions for topk error:analysis andinsights[C]∥2016 IEEEConference on Computer Vision and Pattern Recognition(CVPR),June 27-30,Las Vegas,NV,USA.New York:IEEE,2016:1468-1477.
[21]Russakovsky O,Deng J,Su H,et al.ImageNet large scale visual recognition challenge[J].International Journal of Computer Vision,2015,115(3):211-252.
[2]Zheng H L,Fu J L,Mei T,et al.Learning multiattention convolutional neural network for finegrained image recognition[C]∥2017 IEEEInternational Conference on Computer Vision(ICCV),October 22-29,Venice,Italy.New York:IEEE,2017:5219-5227.
[3]Yan Y,Ni B,Yang X.Fine-grained recognition via attribute-guided attentive feature aggregation[C]∥The 25th ACM International Conference on Multimedia,October 23-27,2017,California,USA.New York:ACM,2017:1032-1040.
[4]Huang K,Zhang B L.Fine-grained vehicle recognition by deep convolutional neural network[C]∥2016 9th International Congress on Image and Signal Processing,Biomedical Engineering and Informatics,October 15-17,Datong,China.New York:IEEE,2016:465-470.
[5]Khusnuliawati H,Fatichah C,Soelaiman R.Multifeature fusion using SIFT and LEBP for finger vein recognition[J].Telecommunication Computing Electronics and Control,2017,15(1):478-485.
[6]Murtaza F,Yousaf M H,Velastin S A.Multi-view human action recognition using 2D motion templates based on MHIs and their HOG description[J].IETComputer Vision,2016,10(7):758-767.
[7]Ling Y G,Hu W P.Vehicle type recognition based on SURF and integral channel features[J].Video Engineering,2016,40(7):139-143.凌永国,胡维平.基于SURF特征与积分通道特征的车型识别[J].电视技术,2016,40(7):139-143.
[8]Fang J,Zhou Y,Yu Y,et al.Fine-grained vehicle model recognition using a coarse-to-fine convolutional neural network architecture[J].IEEE Transactions on Intelligent Transportation Systems,2017,18(7):1782-1792.
[9]Bengio Y,Courville A,Vincent P.Representation learning:A review and new perspectives[J].IEEETransactions on Pattern Analysis and Machine Intelligence,2013,35(8):1798-1828.
[10]Zou Y B,Zhou W L,Chen X Z.Research of laser vision seam detection and tracking system based on depth hierarchical feature[J].Chinese Journal of Lasers,2017,44(4):0402009.邹焱飚,周卫林,陈向志.基于深度分层特征的激光视觉焊缝检测与跟踪系统研究[J].中国激光,2017,44(4):0402009.
[11]Zhou Y C,Xu T Y,Zheng W,et al.Classification and recognition approaches of tomato main organs based on DCNN[J].Transactions of the Chinese Society of Agricultural Engineering,2017,33(15):219-226.周云成,许童羽,郑伟,等.基于深度卷积神经网络的番茄主要器官分类识别方法[J].农业工程学报,2017,33(15):219-226.
[12]Song J,Kim H I,Yong M.Fast and robust face detection based on CNN in wild environment[J].Journal of Korea Multimedia Society,2016,19(8):1310-1319.
[13]Liu B,Yu X C,Zhang P Q,et al.A semi-supervised convolutional neural network for hyperspectral image classification[J].Remote Sensing Letters,2017,8(9):839-848.
[14]Qu L,Wang K R,Chen L L,et al.Fast road detection based on RGBD images and convolutional neural network[J].Acta Optica Sinica,2017,37(10):1010003.曲磊,王康如,陈利利,等.基于RGBD图像和卷积神经网络的快速道路检测[J].光学学报,2017,37(10):1010003.
[15]Cortes C,Vapnik V.Support-vector networks[J].Machine Learning,1995,20(3):273-297.
[16]Elleuch M,Maalej R,Kherallah M.A new design based-SVM of the CNN classifier architecture with dropout for offline Arabic handwritten recognition[J].Procedia Computer Science,2016,80:1712-1723.
[17]Li X Q,Zhang Y,Liao D.Mining key skeleton poses with latent SVM for action recognition[J].Applied Computational Intelligence and Soft Computing,2017,2017:1-11.
[18]Cheng L Y,Mi G Y,Li S,et al.Quality diagnosis of joints in laser brazing based on principal component analysis-support vector machine model[J].Chinese Journal of Lasers,2017,44(3):0302004.程力勇,米高阳,黎硕,等.基于主成分分析-支持向量机模型的激光钎焊接头质量诊断[J].中国激光,2017,44(3):0302004.
[19]Huang F J,LeCun Y.Large-scale learning with SVMand convolutional for generic object categorization[C]∥2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition(CVPR′06),June 17-22,New York,USA.New York:IEEE,2006:284-291.
[20]Lapin M,Hein M,Schiele B.Loss functions for topk error:analysis andinsights[C]∥2016 IEEEConference on Computer Vision and Pattern Recognition(CVPR),June 27-30,Las Vegas,NV,USA.New York:IEEE,2016:1468-1477.
[21]Russakovsky O,Deng J,Su H,et al.ImageNet large scale visual recognition challenge[J].International Journal of Computer Vision,2015,115(3):211-252.