劉玥 孫國(guó)強(qiáng)
摘要:傳統(tǒng)字符識(shí)別方法缺乏對(duì)污染車牌字符正確識(shí)別的能力,難以有效分辨易混淆字符等。針對(duì)這些弊端,采用MATLAB對(duì)真實(shí)車牌字符圖像進(jìn)行處理,提出一種基于離散Hopfield神經(jīng)網(wǎng)絡(luò)的改進(jìn)算法(CLP-HNN),對(duì)車牌字母及數(shù)字進(jìn)行識(shí)別。實(shí)驗(yàn)結(jié)果表明,該算法對(duì)污染車牌字符識(shí)別率達(dá)93.3%,不僅可有效降低污染車牌錯(cuò)誤識(shí)別的風(fēng)險(xiǎn),而且可提高易混淆字符正確辨別率,對(duì)減少車牌誤識(shí)別引起的交通安全及秩序問(wèn)題有較大參考價(jià)值。
關(guān)鍵字:污染車牌;字符識(shí)別;Hopfield神經(jīng)網(wǎng)絡(luò)
DOI:10. 11907/rjdk. 192300 開(kāi)放科學(xué)(資源服務(wù))標(biāo)識(shí)碼(OSID):
中圖分類號(hào):TP301文獻(xiàn)標(biāo)識(shí)碼:A 文章編號(hào):1672-7800(2020)007-0032-04
Contaminated License Plate Character Recognition
Based on Discrete Hopfield Neural Network
LIU Yue, SUN Guo-qiang
(School of Optical-Electrical & Computer Engineering, University of Shanghai for Science & Technology, Shanghai 200093, China)
Abstract: To improve the disadvantages of traditional character recognition methods which lack of ability of correctly recognizing contaminated license plate characters and effectively distinguishing the confusing characters, this paper utilizes MATLAB to process the real license plate character images and proposed the contaminated license plate-Hopfield neural network(CLP-HNN) which is a modified algorithm based on discrete Hopfield neural network to recognize the letters and numbers of contaminated license plate. Experiment results have shown that the recognition rate of contaminated license plate characters by CLP-HNN algorithm can reach 93.3%. It indicates the method proposed in this paper can not only effectively decrease the risk of misrecognition of contaminated license plates but also improve the correct discrimination rate of confusing characters, which is of great significance for reducing traffic safety problems caused by license plate recognition.
Key Words: contaminated license plate; characters recognition; Hopfield neural network
0 引言
智能交通系統(tǒng)(Intelligent Transportation System,ITS)的主要目標(biāo)是在交通運(yùn)輸管理系統(tǒng)中運(yùn)用先進(jìn)的信息、通信、計(jì)算機(jī)等技術(shù)使系統(tǒng)更加實(shí)時(shí)高效[1-2]。車牌識(shí)別技術(shù)作為城市智能交通中采集分析信息的重要方式,承擔(dān)了極其重要的任務(wù)[3-4]。常規(guī)車牌識(shí)別技術(shù)一般分為3個(gè)環(huán)節(jié):定位[5]、分割[6]及識(shí)別[7],環(huán)環(huán)相扣。由于車牌字符正確識(shí)別率直接關(guān)系到車牌識(shí)別系統(tǒng)性能,所以成為完善智能交通管理系統(tǒng)的關(guān)鍵。
然而現(xiàn)實(shí)場(chǎng)景中車牌大多受到程度不一的污染,比如雨雪污泥沾染、人為惡意改動(dòng)以及長(zhǎng)期使用造成的質(zhì)量退化等,這種車牌通常被稱為“污染車牌”,也是當(dāng)前車牌識(shí)別難點(diǎn)之一。大多數(shù)車牌字符識(shí)別是針對(duì)正常車牌的,對(duì)污染字符缺少成熟的手段,無(wú)法確保結(jié)果準(zhǔn)確、高效。因此,如何從這些殘缺、改動(dòng)、模糊的字符中獲取正確完整的字符信息是識(shí)別的關(guān)鍵問(wèn)題。鑒于字母及數(shù)字字符的人為污染可能性及對(duì)識(shí)別結(jié)果的影響程度均大于漢字字符,所以本文主要針對(duì)字母和數(shù)字進(jìn)行研究。
目前常用車牌字符識(shí)別技術(shù)主要分為基于模板匹配的字符識(shí)別算法[8-9]、基于神經(jīng)網(wǎng)絡(luò)的字符識(shí)別算法[10-12]、基于特征統(tǒng)計(jì)匹配法[13]等。文獻(xiàn)[14]提出基于數(shù)學(xué)形態(tài)學(xué)的模糊模板匹配方法,但是對(duì)質(zhì)量差的字符識(shí)別效果欠佳;肖曉等[15]通過(guò)細(xì)化字符字庫(kù),提出一種改進(jìn)的模版匹配算法,在一定程度上克服了傳統(tǒng)模版匹配無(wú)法識(shí)別殘缺字符的缺點(diǎn);Parekh等[16]提出一種新的識(shí)別算法,它以動(dòng)態(tài)生成的車牌字符作為數(shù)據(jù)庫(kù)模板,對(duì)字符進(jìn)行識(shí)別;高強(qiáng)[17]利用張量積小波分解高頻子圖具有方向性的特點(diǎn),提取字符筆畫(huà)特征,得到反映字符結(jié)構(gòu)和統(tǒng)計(jì)特征的聯(lián)和特征向量,從而實(shí)現(xiàn)字符;Masood等[18]詳細(xì)介紹了一種全自動(dòng)車牌檢測(cè)識(shí)別系統(tǒng),該系統(tǒng)核心技術(shù)由深度卷積神經(jīng)網(wǎng)絡(luò)(CNN)等算法結(jié)合而成;Zhang等[19]使用自然圖像訓(xùn)練Hopfield神經(jīng)網(wǎng)絡(luò),以實(shí)現(xiàn)自然圖像的有效壓縮和恢復(fù);Soni等[20]提出一種使用云Hopfield神經(jīng)網(wǎng)絡(luò)識(shí)別低分辨率灰度面部圖像的方法,該網(wǎng)絡(luò)可以處理變形面部,例如戴太陽(yáng)鏡或口罩遮住部分面龐的人。
對(duì)于學(xué)習(xí)率[η],當(dāng)訓(xùn)練樣本為50、訓(xùn)練次數(shù)為80時(shí),學(xué)習(xí)率為0.9,識(shí)別率最高。如表1所示。
對(duì)于訓(xùn)練次數(shù),當(dāng)學(xué)習(xí)率為0.9,訓(xùn)練樣本數(shù)為50時(shí),訓(xùn)練次數(shù)為75和80時(shí)識(shí)別率均比較高,但識(shí)別率為80時(shí),時(shí)延較小,如表2所示。所以本文取學(xué)習(xí)率為0.9,訓(xùn)練次數(shù)為80。
2.3 算法評(píng)估
為驗(yàn)證算法效果,對(duì)算法進(jìn)行綜合對(duì)比:首先對(duì)改進(jìn)的Hopfield神經(jīng)網(wǎng)絡(luò)與傳統(tǒng)Hopfield進(jìn)行縱向?qū)Ρ?然后,將本文算法與其它算法進(jìn)行對(duì)比。
表3中的字符“0”極易認(rèn)為改動(dòng)為“C”、“G”、“Q”、“8”等,“8”易改動(dòng)為“0”等,以這些字符為例展示識(shí)別結(jié)果更具有說(shuō)服力。由表3實(shí)驗(yàn)結(jié)果表明,傳統(tǒng)Hopfield神經(jīng)網(wǎng)絡(luò)不能很好地識(shí)別污染車牌,改進(jìn)的Hopfield神經(jīng)網(wǎng)絡(luò)在識(shí)別結(jié)果上有明顯優(yōu)勢(shì),尤其對(duì)于相似字符本文方法識(shí)別率明顯更高。
不同算法在相同測(cè)試集下的實(shí)驗(yàn)結(jié)果如表4所示。
仿真結(jié)果與實(shí)驗(yàn)數(shù)據(jù)表明,對(duì)于測(cè)試集中的字符識(shí)別率而言,模板匹配算法是最不理想的,由于算法本身特性導(dǎo)致其對(duì)于易混淆字符的識(shí)別錯(cuò)誤率較高;神經(jīng)網(wǎng)絡(luò)算法對(duì)于該類污染字符的識(shí)別更加有效,而本文提出的CLP-HNN算法識(shí)別率最高,污染車牌識(shí)別效果最好。
3 結(jié)語(yǔ)
本文提出一種CLP-HNN算法實(shí)現(xiàn)對(duì)污染車牌字符的識(shí)別,避免了傳統(tǒng)離散Hopfield神經(jīng)網(wǎng)絡(luò)存在的弊端。MATLAB模擬結(jié)果表明,CLP-HNN對(duì)污染車牌的缺失、改動(dòng)及不完整信息有良好的容錯(cuò)性,聯(lián)想記憶成功率也較其它算法更高,識(shí)別結(jié)果更加貼近準(zhǔn)確字符,具有優(yōu)越的污染車牌字符識(shí)別能力。本文實(shí)驗(yàn)僅考慮了數(shù)字和字母字符,尚未驗(yàn)證CLP-HNN算法是否符合漢字識(shí)別,因此將針對(duì)該方向繼續(xù)深入研究。
參考文獻(xiàn):
[1] ZHU L, YU F R. Big data analytics in intelligent transportation systems: a survey [J].? IEEE Transactions on Intelligent Transportation Systems, 2018, 20(1):383-398.
[2] MAIMARIS A, PAPAGEORGIOU G. A review of intelligent transportation systems from a communications technology perspective[C].? Rio de Janeiro: IEEE International Conference on Intelligent Transportation Systems, 2016.
[3] LAN C, LI F, JIN Y. Research on the license plate recognition based on image processing [C].? Qinhuangdao: Fifth International Conference on Instrumentation & Measurement, 2015.
[4] MOLINA-MORENO M,GONZáLEZ-DíAZ I,DíAZ-DE-MARíA F. Efficient scale-adaptive license plate detection system[J]. IEEE Transactions on Intelligent Transportation Systems,2018,(99):1-13.
[5] RAJPUT H, SOM T, KAR S. An automated vehicle license plate recognition system[J]. Computer,2015(48):56-61.
[6] WANG N, ZHU X, ZHANG J. License plate segmentation and recognition of Chinese vehicle based on BPNN[C]. Wuxi: The 12th International Conference on Computational Intelligence and Security,2016.
[7] KHAN M A, SHARIF M, JAVED M Y. License number plate recognition system using entropy-based features selection approach with SVM [J].? IET Image Processing,2018,12(2):200-209.
[8] WANG C M, LIU J H. License plate recognition system [C].? Zhangjiajie: The 2th International Conference on Fuzzy Systems and Knowledge Discovery, 2015.
[9] PURANIC A,DEEPAK K,UMADEVI V. Vehicle number plate recognition system: a literature review and implementation using template matching [J].? International Journal of Computer Applications,2016, 134(1):12-16.
[10] YAO D,ZHU W,CHEN Y. Chinese license plate character recognition based on convolution neural network[C]. Jinan: 2017 Chinese Automation Congress, 2017.
[11] 董峻妃,鄭伯川,楊澤靜. 基于卷積神經(jīng)網(wǎng)絡(luò)的車牌字符識(shí)別[J]. 計(jì)算機(jī)應(yīng)用, 2017, 37(7):2014-2018.
[12] GUAN X Z, ZHANG L. License plate recognition based on improved BP neural network[J]. Techniques of Automation & Applications. 2015, 34(7):66-68.
[13] ZHAO D, SONG J Y, MOUSAVINEZHAD S H. Research on recognition algorithm of free handwritten numerals based on combined structural features [C]. SanDiego: IEEE International Conference on Electro-Information Technology, 2013.
[14] RUAN Z Y,SHEN Y J,LIU F L. An application of mathematical morphology based fuzzy set theory in license plate characters recognition [J].? Computer Engineering & Science,2016,38(3):562-568.
[15] 肖曉,陳杰,邵慧,等. 基于二次模版庫(kù)的車牌殘缺字符識(shí)別[J]. 安徽建筑大學(xué)學(xué)報(bào),2017,25(4):33-37.
[16] PAREKH R, CHAKRABORTY S. An improved template matching algorithm for car license plate recognition[J].? International Journal of Computer Applications, 2015, 118(25):16-22.
[17] 高強(qiáng),劉斌.? 基于提升小波的矩不變量車牌字符識(shí)別方法 [J].? 量子電子學(xué)報(bào), 2016, 33(6):662-670.
[18] MASOOD S Z, SHU G. DEHGHAN A. License plate detection and recognition using deeply learned convolutional neural networks [J].? Computer Vision and Pattern Recognition, 2017.
[19] ZHANG J, ZHUANG T. A novel approach to design weight matrix of Hopfield network[C].? Shanghai: 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference, 2006.
[20] SONI N, SINGH N,KAPOOR A. Face recognition using cloud Hopfield neural network[C]. Chennai:2016 International Conference on Wireless Communications,Signal Processing and Networking, 2016.
[21] HOPFIELD J J. Neural networks and physical systems with emergent collective computational abilities[J]. Proceedings of National Academy of Sciences, 1982, 79(8): 2554-2558.
(責(zé)任編輯:江 艷)