摘要
针对自然场景下街景门牌号码识别困难的问题,提出了一种基于深度网络模型的WMF-CNN(Convolutional neural network with weighted multi-feature fusion,WMF-CNN)模型.该模型采用加权多层特征图融合的思想,首先利用PCA方法对各特征融合图进行降维,然后再根据它们在网络识别过程中的贡献率给予一定的权值,将加权后的图像细节信息与全局逼近信息进行融合,最后将融合特征送入Soft Max分类器,得到识别结果.在国际公开的SVHN数据集上的实验结果表明,所提模型仅需2.2 h便可完成训练,识别率达到95.6%,优于目前的主流算法.此外,所提模型识别单张图片所需的平均时间约为0.38 ms,适用于实时性要求较高的相关应用.
In this paper,a WMF-CNN( Convolutional neural network with weighted multi-feature fusion,WMF-CNN) model based on deep learning is proposed to solve the problem of the recognition on street view house number images in natural scene.The model adopts the idea of weighted multi-layer feature fusion.The PCA method is used to reduce the dimensions of each fusion feature map and then corresponding weights are computed according to their contributions to recognition results.The weighted feature maps representing detailed information are fused with global approximation information provided by the fully connected layer. Finally,the fused features are input to the Soft Max classifier to get a more reasonable recognition result. Our experimental results on SVHN dataset indicate that the proposed WMF-CNN model could be fully trained within 2.2 hours and achieve the recognition rates of 95.6%.Compared with some other methods or models,the suggested WMF-CNN model not only can obtain higher accuracy,but also may meet some the requirements of real-time applications since it takes an average of about 0.38 milliseconds to recognize an image.
引文
[1]PANAHI R,GHOLAMPOUR I.Accurate detection and recognition of dirty vehicle plate numbers for highspeed applications[J].IEEE Transactions on Intelligent Transportation Systems,2017,18(4):767-779.
[2]NGUYEN K,TRAN D,MA W,et al.The impact of data fragment sizes on file type recognition[C].Proceedings of IEEE International Conference on Natural Computation,2014:748-752.
[3]LI D,YU P F,LI H,et al.Printed new tai lue character recognition based on BP neural network[C].Proceedings of IEEE International Conference on Signal and Image Processing,2016:339-342.
[4]KIM D H,KIM J C.An efficient technique for address data input simplification based on the address dictionaries[C].Proceedings of International Conference on Information and Communication Technology Convergence,2016:356-358.
[5]郭健,莫国梁,陈庆伟.基于粒子群优化神经网络的移动机器人门牌识别方法研究[J].计算机与数字工程,2009,37(4):7-9.GUO J,MO G L,CHEN Q W.Research on door number recognition techniques based on PSO-NN for mobile robot[J].Computer&Digital Engineering,2009,37(4):7-9.
[6]张帅,苏世涛.基于形态学走廊环境门牌号的识别[J].现代电子技术,2011,34(14):7-9.ZHANG S,SU S T.Doorplate identification for mobile robot in hall way based on morphological[J].Modern Electronics Technique,2011,34(14):7-9.
[7]马立玲,姬利军,王军政.正交判别的线性局部切空间排列结合SVM的门牌识别[J].中南大学学报:自然科学版,2011,42(S1):789-793.MA L L,JI L J,WANG J Z.An algorithm based on ODLLTSA and SVM classier for door plate number recognition[J].Journal of Central South University:Science and Technology,2011,42(S1):789-793.
[8]GUO Q,TU D,LEI J,et al.Hybrid CNN-HMM model for street view house number recognition[C].Proceedings of Asian Conference on Computer Vision,2014:303-315.
[9]SERMANET P,CHINTALA S,LECUN Y.Convolutional neural networks applied to house numbers digit classification[C].Proceedings of IEEE International Conference on Pattern Recognition,2012:3 288-3 291.
[10]马苗,陈芳,郭敏,等.基于改进Le Net-5的街景门牌号码识别方法[J].云南大学学报:自然科学版,2016,38(2):197-203.MA M,CHEN F,GUO M,et al.A recognition method based on improved Le Net-5 for street view house numbers[J].Journal of Yunnan University:Natural Sciences Edition,2016,38(2):197-203.
[11]LECUN Y,BENGIO Y,HINTON G.Deep learning[J].Nature,2015,521(7553):436-444.
[12]RAMTEKE S P,SHELKE R D,PATIL N P.A neural network approach to printed devanagari character recognition[J].International Journal of Computer Applications,2013,59(16):4 035-4 041.
[13]DU J,WANG Z R,ZHAI J F,et al.Deep neural network based hidden Markov model for offline handwritten Chinese text recognition[C].Proceedings of the International Conference on Pattern Recognition,2016:3 428-3 433.
[14]LECUN Y,BOSER B,DENKER J S,et al.Backpropagation applied to handwritten zip code recognition[J].Neural Computation,1989,1(4):541-551.
[15]QIAN Y,BI M,TAN T,et al.Very deep convolutional neural networks for noise robust speech recognition[J].IEEE Transactions on Audio Speech&Language Processing,2016,24(12):2 263-2 276.
[16]ALBELWI S,MAHMOOD A.Automated optimal architecture of deep convolutional neural networks for image recognition[C].Proceedings of IEEE International Conference on Machine Learning and Applications,2017:53-60.
[17]NETZER Y,WANG T,COATES A,et al.Reading digits in natural images with unsupervised feature learning[C].NIPS Workshop on Deep Learning and Unsupervised Feature Learning,2011(2):1-9.