人臉識別外文翻譯---對于臉部識別系統(tǒng)研究遺傳算法

1、<p><b>　　外文科技資料翻譯</b></p><p><b>　　英文原文</b></p><p>　　Research on a face recognition system by the genetic algorithm</p><p>　　Computer vision and recognit

2、ion is playing an increasingly important role in modern intelligent control.Object detection is the first and most important step in object recognition. Traditionally,a special object can be recognized by the template ma

3、tching method,but the recognition speed has always been a problem.In this article,an improved general genetic algorithm-based face recognition system is proposed.The genetic algorithm(GA)has been considered to be a robus

4、t and global searching method.He</p><p>　　1 Introduction</p><p>　　If we search on the web or in a conference proceedings about intelligent control,lots of papers and applications are presented.A

5、mong them,image processing and recognition occupy a very large percentage.The higher the degree of intelligence,the more important the image detection and recognition technology.</p><p>　　For controlling an

6、intelligent system(autonomous mobile vehicle,robot,etc.),the most important element is the control strategy,but before automatically making it move,image recognition is needed.For an intelligent control system,it is nece

7、ssary to acquire information about the external world automatically by sensors,in order to recognize its position and the surrounding situation.A camera is one of the most important sensors for computer vision.That is to

8、 say,the system endeavors to find out wha</p><p>　　The reliability and time-response of object detection and recognition have a major influence on the performance and usability of the whole object recognitio

9、n system.1The templatematching method is a practicable and reasonablemethod for object detection.2This article gives an improvement in the general template matching method.</p><p>　　In addition,in order to s

10、earch for the object of interest in an image,lots of data need to be processed.The geneticalgorithm(GA)has been considered to be a robust and global searching method(although it is sometimes said thatGA can not be used f

11、or finding the global optimization3). Here,the chromosomes generated by GA contain information about the image data,and the genetic and evolution operations are used to obtain the best match to the template:searching for

12、 the best match is the goal of this</p><p>　　In this article,Sect.2 gives the encoding method of the GA and the experimental setting that is used.In Sect.3,the experiment and the analysis are addressed.Some

13、conclusions are given in Sect.4.</p><p>　　Theory and experimental setting</p><p>　　For an image recognition system,the most interesting part that has special features has first to be detected in

14、 the original image.This is called object detection.After that,this part will be compared to a template to see if it is similar or not.This is called object recognition.For example, if we want to find a special person in

15、 an image,we first have to detect people in the image,and then recognize which one is the person of interest(sometimes these two steps will be executed simultaneously).T</p><p>　　Fig.1. Object recognition sy

16、stem</p><p>　　Statistical object recognition involves locating and isolating the targets in an image,and then identifying them by statistical decision theory.One of the oldest techniques of pattern recogniti

17、on is matching filtering,4which allows the computation of a measure of the similarity between the original image f(x,y)and a template h(x,y).Define the mean-squared distance</p><p><b>　　(1)</b>&l

18、t;/p><p>　　,if the image and template are normalized by</p><p><b>　　(2)</b></p><p><b>　　And then</b></p><p><b>　　=</b></p><p

19、>　　= (3)</p><p>　　For the right-hand side of Eq.3,the first term is constant,and thus can measure as the least-squared similarity between the original image an

20、d the template.5 If has alarge value (which means that is small enough),then the image is judged to match the template.If is less than apreselected threshold,the recognition process will either reject the match or cre

21、ate a new pattern,which means that the similarity between the object in the original image and the template is not satisfied.</p><p>　　2.1 Genetic encoding</p><p>　　As introduced above,the chrom

22、osomes generated by the GA contain information about the image data,so the first step is to encode the image data into a binary string.6The parameters of the center of a face(x,y)in the original image,the rate of scale t

23、o satisfy eq.2,and the rotating</p><p>　　angle θ are encoded into the elements of a gene.Some important parameters of the GA used here are given in Table 1,and the search field and region are given in Table2

24、.As shown in Table 2,one chromosome contains 4 bytes:the coordinate(x,y)in the original image,the rate of scale,</p><p>　　and the rotation angle θ.</p><p>　　2.2 Experimental setting</p>&

25、lt;p>　　The experiment is done by first loading the original and the template images.The GA is used to find whether or not there is the object of a template in the original image.If the answer is YES,then in the origin

26、al image the result gives the coordinates of the center of the object,the scale,and the rotation angle from the template.For comparison,the general template matching method is also presented.7The execution time shows the

27、 effectiveness of the GA-</p><p>　　based recognition method. Figures 2 and 3 are the original images and the templates for the experiment.The values are the width×height in pixels of the image.In Fig.2,

28、three images are presented,the content and size of which are different.Figure 2a has two faces(the faces of a person and a toy),Fig.2b shows a face tipped to one side,and the person in Fig.2c wears a hat</p><p

29、>　　and the background is more complicated than in Figs.2a and b.</p><p>　　The two templates in Fig.3 are not extracted from the same image.For normal use,the template should be extracted as the average of

30、 several feature images.In Fig.4,the template(a)-0 is generated from(a)-1,(a)-2,and(a)-3,and takes the average value of the gray levels from the three models.The same is also true for(b)-0.</p><p>　　Fig.2.Th

31、ree original images(max_x×max_y).a 238×170.b 185×196.c 275×225</p><p>　　Fig.3.Templates for matching(temp_x×temp_y).a Template 1.b Template 2</p><p>　　3 Experiment and c

32、omparison</p><p>　　The genetic operations and GA parameters are presented in Table 1 and Table 2.The fitness is defined as</p><p><b>　　(4)</b></p><p>　　In Eq.4, is the g

33、ray level of the coordinatesin the template image,the width and height of which are and .gives the gray level in the original image,the coordinates of which are calculated by translation from,and by changing the scale a

34、nd the rotation angleθfrom the template.Since the images are 256 gray-level images,in Eq.4,division by 255 ensures that the resulting fitness is between 0 and 1.The maximum number of generations is limited to 300,and the

35、 threshold of the matching rate is set to 0</p><p>　　The results of GA-based face recognition are given in Fig.6 and Table 3.Figure 6a,c and d are searched to match the template Fig.3a,while Fig.6b is matche

36、d to Fig.3b. Figure 6a and b reach the matching rate 0.9 within 300 generations,while Fig.6c and d cannot reach the matching rate 0.9 within 300 generations(the best match is given in Table 3).In the images in Fig.6a–c,w

37、e see that the result given matches the template well.The coordinates ,the rate of scale,and the angle of rotationθhave been f</p><p>　　Fig.4.Creation of template</p><p>　　For the purpose of com

38、paring the effects of the GA-based algorithm,the result of the general matching method7 is also presented.From Fig.5,we see that although both the original image(the top-left image)and the template(the top-right image)ar

39、e simplified by binarization,the matching time is 1 min 22 s.The recognizable result is the bottomleft image in Fig.5.</p><p>　　Fig.5.Result of searching by a GA</p><p>　　4 Conclusions</p>

40、<p>　　In this article,the GA-based image recognition method is tested,and a comparison with the general matching method is presented.</p><p>　　As we know,the GA starts with an initial set of random so

41、lutions called the population.Each individual in the population is called a chromosome,and represents a solution to the problem.By stochastic search techniques based on the mechanism of natural selection and natural gene

42、tics, genetic operations(crossover and mutation)and evolutionary operations(selecting or rejecting) are used to search for the best solution.8</p><p>　　In this article,the chromosomes generated by the GA con

43、tain information about the image, and we use the genetic operators to obtain the best match between the original image and the template.The parameters are the coordinates(x,y)of the center of the object in the original i

44、mage, the rate of scale,and the angle of rotationθ.</p><p>　　In fact,translation,scale,and rotation are the three main invariant moments in the field of pattern recognition.9However,for face recognition,the

45、facial features are difficult to extract,and are calculated by the general pattern recognition theory and method.10 Even these three main invariant moments will not be invariant because the facial expression is changed i

46、n different images.</p><p>　　Thus,recognition only gives the best matching result within an upper predetermined threshold. Both the GA-based method and the general template matching method are presented here

47、,and the comparison with the traditionalpattern matching method shows that the recognition is satisfactory,although under some conditions the result is not very good(Fig.6d).</p><p>　　Based on the results of

48、 the experiments described here,future work will emphasize(i) optimizing the fields of chromosomes,and(ii)improving the fitness function by adding some terms to it.This work is important and necessary in order to improve

49、 the GA-based face recognition system.</p><p>　　References</p><p>　　1.Sugisaka M,Fan X(2002)Development of a face recognition system for the life robot.Proceedings of the 7th International Sympo

50、sium on Artificial Life and Robotics,Oita,Japan,vol 2,Shubundo Insatsu Co.Ltd.,pp 538–541</p><p>　　2.Castleman K(1998)Digital image processing.Original edition published by Prentice Hall;a Simon&Schuster

51、 Press of Tsinghua University,China</p><p>　　3.Iba H(1994)Foundation of genetic algorithm:solution of mystic GA(in Japanese).Omu Press</p><p>　　4.Deguchi K,Takahashi I(1999)Image-based simultane

52、ous control of robot and target object motion by direct-image interpretation.Proceedings of the 1999 IEEE/RSJ International Conference on Intelligent Robot and Systems,Kyongju,Korea,vol 1,pp 375–380</p><p>　

53、　5.Jaehne B(1995)Digital image processing:concepts,algorithms,and scientific applications,3rd edn.Springer Berlin,Heidelberg,Germany</p><p>　　6.Agui T,Nagao T(2000)Introduction to image processing using prog

54、ramming language C(in Japanese).Shoko-do Press</p><p>　　7.Ishibashi’s studying room of C++(in Japanese).http://homepage3.nifty.com/ishidate/vcpp.htm</p><p>　　8.Gen M,Cheng R(1997)Genetic algorit

55、hms and engineering design.Wiley-Interscience,New York</p><p>　　9.Agui T,Nagao T(1992)Image processing and recognition(in Japanese).Syokoudou Press</p><p>　　10.Takimoto H,Mitsukura T,Fukumi M,et

56、 al.(2002)A design of a face detection system based on the feature extraction method.Proceedings of the 12th Symposium on Fuzzy,Artificial Intelligence,Neural Networks and Computational Intelligence,Saga,Japan,pp 409–410

57、</p><p><b>　　中文譯文</b></p><p>　　對于臉部識別系統(tǒng)研究遺傳算法</p><p>　　基于計算機視覺的手勢識別對于當今智能控制起著非常重要的作用。通常，模板圖像匹配法能夠識別一個特殊物體，但是識別的速度總是個問題。這篇文章是討論一個改進的人臉識別遺傳算法系統(tǒng)。遺傳算法已經(jīng)被認為是一個穩(wěn)健的全局搜索的方法。在這，G

58、A產(chǎn)生的染色體含有識別這個物體的信息。這篇文章提出了一個人臉識別的有效方法。最后，列舉了各種實驗結(jié)論，和傳統(tǒng)的模板圖像匹配法對比，還有一些其他考慮因素。</p><p><b>　　引言</b></p><p>　　如果我們搜索關(guān)于智能控制的網(wǎng)站或者回憶進程，我們就能找到很多文章或者應用文。這些文章大部分是關(guān)于圖像處理和圖像識別。智能度越高，圖像處理和識別的技術(shù)越高。

59、對于控制和智能系統(tǒng)來說（自動移動工具，機器人等），最重要的因素是控制策略，但是在是它們自動運作之前，圖像識別也是必要的。對于智能控制系統(tǒng)來說，通過感應器自動獲得外部世界的信息比如識別它的位置和周圍的環(huán)境也是很有必要的，照相機是計算機視覺的一個重要感應器。也就是說，這個系統(tǒng)能夠確認照相機（或者機器人）攝入的圖像：交通信號，障礙物，還有指南等</p><p>　　這個物體是別的可靠性和感應速度對于其運作和左右有著非常

60、重要的作用。模板圖像匹配法是一種更有效且合理的物體識別方法。這篇文章主要是詳述普通模板圖像匹配法。除此之外，需要很多數(shù)據(jù)才能搜索到圖像上的土體。基因遺傳算法被認為是一種穩(wěn)健的全球搜索方法（盡管有時基因遺傳算法不能用于全球優(yōu)化。這里，遺傳基因算法所得的染色體含有各種信息關(guān)于圖像和數(shù)據(jù)，以及基因和演化操作能夠傳統(tǒng)模幫助獲取板圖像匹配：尋找最好的匹配數(shù)據(jù)是這篇文章的目標。這個想法來源于GA的特征，以及通過智能系統(tǒng)快速簡單地識別人臉的需要。這里

61、不僅僅是介紹這個圖像處理和GA的概念，因為，這個話題已經(jīng)有更詳細的文章論述了。</p><p>　　在這篇文章，第二部分介紹GA的解碼方法還有所需的實驗設(shè)備。第三部分主要是陳述實驗和分析。第四部分主要是一些結(jié)論。</p><p><b>　　理論和實驗設(shè)備</b></p><p>　　對于影響識別系統(tǒng)來說，最有趣的部分是在原始圖像中已經(jīng)首先發(fā)現(xiàn)

62、了特殊的特征。這就被稱為物體識別。其次，這一部分會和模板圖像對比以確認他們相似否。這就被稱為目標檢測。例如，如果我們想在一張圖像中找到一個人特殊特征，我們首先應該要在這幅圖像中找到這個人，然后識別出哪個是這個人的特征（有時候這個兩個步驟會有序進行）這一個步驟在圖一有所展示。</p><p><b>　　圖1物體識別系統(tǒng)</b></p><p>　　統(tǒng)計識別函數(shù)包括定位

63、和分離圖像中的兩個目標，然后通過統(tǒng)計確認理論識別他們。這個最老的系統(tǒng)識別技術(shù)是匹配分離，計算方法可以獲得原始圖像(x,y）和模板h(x.y),的相似之處。最后定義均方距離。</p><p><b>　　(1)</b></p><p>　　,對圖像和模板進行歸一化</p><p><b>　　(2)</b></p>

64、;<p><b>　　然后得到</b></p><p><b>　　=</b></p><p>　　= (3)</p><p>　　為了公式（3）右邊，在第一個周期內(nèi)是恒定的，因此可以作為衡量與原始圖像和模板最小平方的相

65、似性。[5] 如果有一個較大的值（即足夠小），則判定該圖像匹配模板。如果比預選門檻低，識別過程要么拒絕匹配要么創(chuàng)建一個新的模式，這意味著在原來的圖像相似性的對象和模板不滿意。2.1基因破解</p><p>　　綜上所述，GA的染色體含有圖像數(shù)據(jù)的信息，所以第一步是將圖像數(shù)據(jù)分解為二進制串位。在原圖像中臉中心的周長，滿足公式2的比例還有旋轉(zhuǎn)角被分解為一個基因的多個因素。這里GA一些重要的周長以表1的形式展示出來了

66、。圖表2是搜索范圍和地區(qū)。正如圖表2所展示的，一個染色體含有四個字節(jié)：坐標，原圖像還有比例，以及旋轉(zhuǎn)角。</p><p><b>　　2.2實驗設(shè)置</b></p><p>　　這個實驗是由第一加載原和模板圖像應用遺傳算法找到了是否存在的目的是一個模板,在原圖像的。如果答案是肯定的,那么在原始圖像結(jié)果給出坐標中心的對象、規(guī)模、旋轉(zhuǎn)角度從模板與之相比,一般的模板匹配法提

67、出了執(zhí)行時間顯示。基于遺傳算法的有效性的識別方法。</p><p>　　如圖2和3是原始圖像和模板進行實驗,這個值×寬的圖像像素高度…在圖2、3圖像和內(nèi)容,提出了具有不同大小圖2有兩副面孔的人的臉,一個玩具),圖2中的b顯示一臉向一邊,并將在圖2中的c的人戴著一頂帽子更復雜的背景下,在圖2中的(a)和(b)。</p><p>　　這兩個樣板圖不是提取相同的圖像正常使用模板應提取的

68、幾個特征的平均圖像圖、模板(a)-0是來自(a)-1,(a)-2、(a)-3,并以平均價值的灰度層次的三種模式的. 同樣的情況也英尺(b)-0。</p><p>　　圖2 三張原來圖像(max_x×max_y) 238×170.b 185×196.c 275×225</p><p>　　圖3 模板匹配(temp_x×temp_y模板1.

69、b模板)</p><p>　　三，比較試驗遺傳操作,給出了算法參數(shù)表1和表2。 健康的定義是</p><p><b>　　(4)</b></p><p>　　在公式4, 在模板圖像中，是的水平坐標系數(shù)，坐標系的寬度分別為和高度。提供了原始圖像灰度層次、坐標的計算,并通過轉(zhuǎn)化可以通過改變規(guī)模和旋轉(zhuǎn)角度，在公式（4）中，除以255確保所產(chǎn)生的結(jié)果

70、是在0和1之間，最多的后代限于300的門檻。匹配率確切地說，如果在3代匹配率可達0.9，那么說明模板中發(fā)現(xiàn)了原始圖像(模板匹配的原始圖像閾值)。否則，結(jié)果要訓練達到300代。</p><p>　　圖6和表3 給出了基于遺傳算法的人臉識別結(jié)果。圖6( a)、(c)、(d)搜索匹配的模板匹配,而圖6(b )匹配圖3（b）。圖6（a)和(b)匹配率達到300世代之內(nèi),而圖6中c和d不能達到300代以內(nèi)，在圖像中最佳匹配

71、率0.9的 (表3)是我們看到了圖6a-c。結(jié)果給出了模板匹配坐標、規(guī)模化、旋轉(zhuǎn)匹配圖6（d）正確，但結(jié)果并不十分滿意的。 其原因是這個模板圖3（a）不能代表在任何時候都有用，也就是說，盡管人在不同的圖像被認出來是相同的，但是這個模板在不同的外觀等所有條件不能給出這個人的特點。另一個原因是，該算法本身也存在著一些問題。例如，用了一種基于遺傳算法的識別方法，設(shè)置搜索的字段(本文中，被選中)，應該確定的遺傳操作，同時選擇和優(yōu)化適應度的函數(shù)。

72、</p><p><b>　　圖4 創(chuàng)建模板</b></p><p>　　為了比較的基于遺傳算法的算法的影響，一般匹配方法[7]結(jié)果的目的。從圖5，我們看到，雖然無論是原圖像（左上圖）和模板（右上方圖片）是由二值化簡化，匹配時間為1分22 s。辨認的結(jié)果是在圖5 底部左邊的形象。</p><p>　　圖5由遺傳的搜索結(jié)果</p>

73、<p><b>　　結(jié) 論</b></p><p>　　本文中研究了基于遺傳算法的圖像識別方法，并且給出了和一般圖像匹配算法的比較結(jié)果。</p><p>　　我們都知道遺傳算法開始于一組初始的隨機解集，稱之為種群。種群中的每一個個體叫做染色體，其代表了問題的一個解。通過模仿自然選擇和遺傳機制進行隨機搜索，使用基因操作（交叉和變異）和演化操作（選擇或者遺棄）

74、來搜索最優(yōu)解。</p><p>　　本文中，遺傳算法產(chǎn)生的染色體包含了圖像信息，我們可以通過遺傳算子得到原始圖像和模板之間的最優(yōu)匹配。參數(shù)包括原始圖像中物體的中心點坐標（x，y），尺度變換率以及旋轉(zhuǎn)角度θ。</p><p>　　事實上，平移、尺度變化和旋轉(zhuǎn)是模式識別中的三個主要不變矩。然而，對于臉部識別來時，使用一般的模式識別的理論和方法進行臉部特征的提取和計算比較困難。甚至由于在不同的圖

75、像中臉部特征的表達是會改變的，這三個主要的不變矩將不再是不變的。</p><p>　　所以，識別僅僅是在一個預先確定的上限閾值內(nèi)給出了最優(yōu)匹配。本文同時給出了基于遺傳算法的方法和一般的模板匹配方法，和傳統(tǒng)的模式識別方法的比較表明這種識別方法盡管在某些情況下其結(jié)果并不是很理想（圖6d），但是其結(jié)果總體上還是比較令人滿意的。</p><p>　　基于以上實驗的結(jié)果，我們未來的工作主要集中在以下