《快速人脸检测技术综述.pdf》由会员分享，可在线阅读，更多相关《快速人脸检测技术综述.pdf（18页珍藏版）》请在三一文库上搜索。
1、 121,2,3 1 1 100022 2 150001 3 100080 2090 21ViolaAdaBoost Boosting TP391.4 Face Detection: a Survey Yuemin Li1 Jie Chen2 Wen Gao1,2,3 Baocai Yin1 1(Multimedia and Intelligent Software Technology Laboratory Beijing University of Technology, Beijing 100022, China) 2(School of Computer Science and Tec
2、hnology, Harbin Institute of Technology, Harbin, 150001, China) 3 (Institute of Computing Technology, Chinese Academy of Sciences, Beijing, 100080, China) , , , Abstract: Face detection is born, as an independent subject, of face recognition and develops with the requirement of the automatic face r
3、ecognition system. Over the past ten years face detection has been thoroughly studied in computer vision research for its interesting applications, such as a face recognition system, a surveillance system and a machine interface. Whether face detection can come into use depends on the two key proble
4、ms: the detection rate and the velocity together. Although the detection rate has been improving during the last ten years, the speed is still a problem to cumber face detection system from being widely used. The hard work of researchers, especially the release of the AdaBoost based classifier prese
5、nted by Viola, has made the detection velocity improve rapidly. Since, the researches began to lay more emphasis on the velocity of the system and a lot of algorithms, about how to increase the speed of face detection, have been presented. Based on these rapid developments of its velocity, this pape
6、r demonstrates it from the view of velocity. The whole stage of face detection is divided into four parts according to the extent of the face-detection velocity: the initial phase, the developing phase, the turning point and the synthesis phase. After the systematical analyses of the papers in diffe
7、rent phases, several promising directions for future researches are also proposed in this paper. Key words: Face detection, velocity, face recognition, pattern recognition, Boosting 1 1,2 (face recognition)(face detection) 3 4, 5 face detection 6 (1) (2) (1) (2) (3) 6,7,8 910,11 2 MITSung12k- 6(Clus
8、ters)6 1 Sung(bootstrap) ( ) 10,11,13 Fig. 1. Face and nonface clusters used by Sung and Poggio 12. Bottom row are the final models which consist of six Gaussian “face” clusters and six “nonface” clusters 12( Courtesy of Tomao Poggio). 1. Sung Poggio M. H. YangSNoW (Sparse Network of Winnows) 13SNoW
9、14, 15Winnow16 102,400 Olivetti, UMIST, Harvard, Yale FERET 17, Kullback 18, Bayes19SVM20 H. Schneiderman 1,19 1717 2 3. 3.1 3.1.1. Michael J. Jones 21B. MartinkauppiSkin locus 22 Fig. 2. Skin detection: (a) a yellow-based face image; (b) skin regions of (a) shown in white; (c) a lighting compensate
10、d image of (a); (d). Skin regions of (c) 27( Courtesy of R. L. Hsu) 2. (a) (b) (a) (c)(d) (c). 23 24skin locus Dorin Comaniciu25, 26 mean shift R. L. Hsu 272 SobottkaPitas28 HSV J. C. Terrillon 29, 30 FourierMellin11 100 85%31, 32, 3.1.2 H.Kim 33Wang 34 B. FrbaEOM (Edge-Orientation Matching)35, 36 E
11、OM3 37 38 Fig. 3. Example of an edge orientation vector field 36( Courtesy of Bernhard Frba). 3. J.Miao 3940, 41, 42, 3.2 B. Frba10 11,43 Craw Sobel 44H. Schneiderman 9 45,46 47 Fig.4. Example of the hierarchical grid search 10. 4. 3.3 multi-thread Shpungin48 3.4 C. J. LiuBayes (BDF) 49 Haar5 Bayes
12、161631616=768 LiuPCA1076.8 Bernd Heisele45283PCA 1006030 3.5 39,50Rowley g2 g1 g0 (a) (c) (b) (d) Fig. 5. Discriminating feature analysis of the mean face and the mean nonface. (a) The mean face and the two bar graphs are its amplitude projections; (b) 1D Harr wavelet representation of the mean face
13、; (c) The mean nonface and its amplitude projections; (d) 1D Harr wavelet representation of the mean nonface 65. (Courtesy of R. Lienhart ) 5. (a) (b) Harr(c) (d) Harr 516 -360o, 360o 17.650 Fig. 6: Overview of the algorithm 51. (Courtesy of H. Rowley ) 6. 52 Juell 53Kou Zani54Anifantis 55 3.6 C. Sa
14、nderson 56 Edgar OsunaSVM59 Jie Yang 58 Kazumasa Murai59 Purdy Ho 60 Raphael Feraud61 Ying Zhu 62 4. P. Viola 11 AdaboostCascade 63 17 (x, y) , ( , )( xx yy ii x yi x y ( , )ii x y( , )x y( , )i x y( , )ii x y ( , ) ( ,1)( , )s x ys x yi x y( , ) ( -1, )( , )ii x yii xys x y( , )s x y ( , -1)0s x(-1
15、, )0iiy Fig. 7. Calculation of the integral image value: The value of the integral image at location 1 is to calculate the sum of the pixels in rectangle A. The value at location 2 is A+B, at location 3 is A+C, and at location 4 is A+B+C+D. The sum within D can be computed as 4+1-(2+3) 11. (Courtesy
16、 of P. Viola) 7. 1A2 A+B3A+C4A+B+C+DD4+1-(2+3) 78 Haar-Like8 Fig. 8. Example rectangle features demonstrated in the enclosing detection window. The pixels sum within the white rectangles is subtracted from the pixels sum in the grey rectangles. Two-rectangle features are shown in (A) and (B). Figure
17、 (C) shows a three-rectangle feature, and (D) a four-rectangle feature 11. (Courtesy of P. Viola ) 8. (A)(B)Harr-like(C) Harr-like(D) Harr-like 2AdaBoost AdaBoost Viola180,000 ( ) j h x j f jj p: 1if ( ) ( ) 0otherwise jjjj j p fxp h x x24 24 1 Table 1. The Adaboost algorithm of classifier learning
18、1 Adaboost 11 Given example set S and their initial weights 1; Do for t=1,T: 1 Normalize the weights t; 2 For each feature, j, train a classifier hj with respect to the weighted samples; 3 Calculate error, choose the classifier ht with the lowest error; 4 Update weights 1t ; Get the final strong cla
19、ssifier h(x). Fig. 9. The overview of the detection cascade. Each sub-window is scanned by a series of classifiers. A large number of negative examples are eliminated by the first classifier with very little processing. Subsequent layers eliminate additional negatives. The number of sub-windows has
20、been reduced radically after several stages of processing. 11. (Courtesy of P. Viola ) 9. 15 5 P. ViolaAdaBoost Boosting 1.(Edge features) 2(Line features) 3 (Center-surround features) 4 (Special diagonal line) feature used Fig. 10. Feature prototypes of simple haar-like and center-surround features
21、. Black areas have negative and white areas positive weights 43( Courtesy of Rainer Lienhart). 10. Harr-like Rainer LienhartViolaHaar-like1043, 64, 65 10%Boosting(Discrete Adaboost) (Real Adaboost) (Gentle Adaboost)(Gentle Adaboost) B. FrobaAdaboost 10 EOM(edge orientation matching) 36SNoW Athlon100
22、0MHz3842880.05 Stan Z. LiFloatBoost 66, 67 FloatBoost AdaBoost Pentium-III700 MHz CPU320240200ms 68 Fig. 11. Face features found by Kullback-Leibler Analyses (KLA). The first row lists some features sequentially found in KLBoosting. The second, third and last rows are “global semantic”, “global but
23、not semantic” and “local” features, respectively 69. (Courtesy of C. Liu ) 11. K-L1KLBoosting23 C. LiuKullback-Leibler Boosting (KLB) (compact) 69 AdaBoost KL KLKL11 CMU P. ViolaAdaboost 70 AdaboostZhang Boosting71SahbiSVM 72,73 , CMU74 2 2 8 Table 2. Some parameters of different methods (appear acc
24、ording to the time of being printed) 2 Reference Methods Velocity (frame/second) Detection rate (%) False alarms Image size (by pixels) Configuration of computer The year of being printed H. Rowley 171 86.0 31 320240 1998 H. Schneiderman 90.2 94.4 65 320240 2000 Y. Zhu 751 No performance improvement
25、 over 75 320240 PC 300MHz 2000 R. Feraud 611 74.7 46 108108 10241024 DEC Alpha 333MHz 2001 P. Viola 1115 92.1 50 384288 PIII700MHz 2001 R. Lienhart 435 82.3 24 320240 P4 2GHz 2002 Stan Z. Li 675 90.2 31 320240 PIII 700MHz 2002 C.J. Liu 49 1 97.4 1 320240 900MHz Sun Blade 1000 workstation 2003 C. Liu
26、 692.5 95.0 10-6 320240 P4, 1.8GHz 2003 B. Frba 1025 89.7 22 320240 Athlon1000MHz 2003 Denote that the detector can run at 1frame per second on the images of 320240, approximately; It is the average velocity while the detector runs on the images ranging from 108108 to 10241024; Only the false rate i
27、s given instead of the false alarms; It is acquired by the ROC curves in this paper. 3202401 13,182108108-10241024 ROC 6 BoostingCascade Haar-likeBoosting 4 AdaboostAdaboost 98.5% Bernhard Froba10 76 77 9, 10, 66, 67, 683D 3D 80% 7. 863(2001AA114190, 2002AA118010) “ ”(60332010)( KP0706200377)“973”(2
28、001CCA03300)“863” (2001AA114160)(D070601-01) (P070701-01,P070702-01) Reference: 1 Chellappa RWilson CLSirohey S. Human and machine recognition of facesA survey. Proceedings of the IEEE199583(5)705740. 2 Zhou JLu CYZhang CS. A survey of automatic face recognition. ACTA Electronica Sinica200028(4)1021
29、06(in Chinese with English abstract). 3 Samaria FS. Face Recognition Using Hidden Markov Models. PhD thesis, Univ. of Cambridge, 1994. 4 Belhumeur P, Hespanha J, Kriegman D. Eigenfaces vs. Fisher faces: Recognition Using Class Specific Linear Projection. IEEE Trans. Pattern Analysis and Machine Inte
30、lligence, vol. 19, no. 7, pp. 711720, 1997. 5 Lu CY. Research on some issues of face recognition and system implementation PhD dissertation. Tsinghua University, Beijing, 1998(in Chinese with English abstract). 6 Yang MH, Kriegman D, Ahuja N. Detecting faces in images: A survey. IEEE Trans Pattern A
31、nalysis and Machine Intelligence, 2002, 24 (1): 3458. 7 Hjelmas E. Face Detection: A Survey. Computer Vision and Image Understanding 83,236274(2001). 8 Liang LH, AI HZ, Xu Z, Zhang B. A Survey of Computer Science. Chinese Journal of Computer,2002,25(5):527532(in Chinese with English abstract). 9 Sch
32、neiderman H, Kanade T. Probabilistic modeling of local appearance and spatial relationships for object recognition. In: Proc IEEE Conference on Computer Vision and Pattern Recognition, Santa Barbara, California, 1998.4551. 10 Frba B, A. Ernst. Fast Frontal-View Face Detection Using a Multi-Path Deci
33、sion Tree. In Proc. Audio- and Video-based Biometric Person Authentication (AVBPA 2003), pp. 921928, 2003. 11 Viola P. Rapid object detection using a Boosted cascade of simple features. In: Proc IEEE Conference on Computer Vision and Pattern Recognition, pp:511518, 2001. 12 Sung KPoggio T. Example-b
34、ased learning for view based human face detection. IEEE Trans Pattern Analysis and Machine Intelligence199820(1)3951. 13 Yang MH, Roth D, Ahuja N. A SNoW-Based Face Detector. Advances in Neural Information Processing Systems 12, S.A. Solla, T. K. Leen, and K.-R. Muller, eds., pp. 855861, MIT Press,
35、2000. 14 Rot D. Learning to Resolve Natural Language Ambiguities: A Unified Approach. Proc. 15th Natl Conf. Artificial Intelligence, pp. 806813, 1998. 15 Carleson A, Cumby C, Rosen J, and Roth D. The SNoW Learning Architecture. Technical Report UIUCDCS-R-99-2101, Univ. of Illinois at Urbana-Champaig
36、n Computer Science Dept., 1999. 16 Littlestone N. Learning Quickly when Irrelevant Attributes Abound: A New Linear-Threshold Algorithm. Machine Learning, vol. 2, pp. 285318, 1988. 17 Rowley H, Baluja S, and Kanade T. Neural Network-Based Face Detection. IEEE Trans. Pattern Analysis and Machine Intel
37、ligence, vol. 20, no. 1, pp. 2338, Jan. 1998. 18 Colmenarez AJ, Huang TS. Face Detection with Information-Based Maximum Discrimination. IEEE Conf. Computer Vision and Pattern Recognition. pp. 782787, 1997. 19 Schneiderman H. A statistical method for 3D object detection applied to faces and cars. In
38、International Conference on Computer Vision, 2000 20 Osuna E, Freund R, and Girosi F, Training Support Vector Machines: An Application to Face Detection. Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp. 130136, 1997. 21 Jones MJ,. Rehg JM. Statistical color models with application to sk
39、in detection. Technical report, Cambridge Res. Lab., Compaq Computer Corp., 1998. 22 Martinkauppi B. Face colour under varying illumination-analysis and applications. PhD thesis , University of Oulu, 2002. 23 Terrillon JC, Shirazi M N, Fukamachi H, and Akamatsu S. Comparative performance of differen
40、t skin chrominance models and chrominance spaces for the automatic detection of human faces in color images. Proc. IEEE Intl Conf. on Face and Gesture Recognition, pp. 54-61, 2000. 24 Hadid A, Pietikainen M, Martinkauppi B. Color-based face detection using skin locus model and hierarchical filtering
41、, ICPR02 (IV: 196200). 25 Comaniciu D, Ramesh V, Meer P. Real-Time Tracking of Non-Rigid Objects using Mean Shift. IEEE Conf. Computer Vision and Pattern Recognition (CVPR00), Hilton Head Island, South Carolina, Vol. 2, 142149, 2000. 26 Comaniciu D, Ramesh V. Robust Detection and Tracking of Human F
42、aces with an Active Camera. IEEE Int. Workshop on Visual Surveillance, Dublin, Ireland, 1118, 2000. 27 Hsu RL, Abdel-Mottaleb M, A K Jain. Face detection in color images. IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 24, no. 5, pp. 696706, May 2002. 28 Sobottka K and Pitas I. Face Loca
43、lization and Feature Extraction Based on Shape and Color Information. Proc. IEEE Intl Conf. Image Processing, pp. 483486, 1996. 29 Terrillon JC, David M,. Akamatsu S. Automatic Detection of Human Faces in Natural Scene Images by Use of a Skin Color Model and Invariant Moments. Proc. Third Intl Conf.
44、 Automatic Face and Gesture Recognition, pp. 112117, 1998. 30 Terrillon JC, David M, Akamatsu S. Detection of Human Faces in Complex Scene Images by Use of a Skin Color Model and Invariant Fourier-Mellin Moments. Proc. Intl Conf. Pattern Recognition, pp. 13501355, 1998. 31 Garcia C, Tziritas G. Face
45、 detection using quantized skin color regions merging and wavelet packet analysis. IEEE Trans Multimedia, 1999,1(3):264277. 32 Yang MH, Ahuja N. Detecting human faces in color images. In: Proc IEEE Conference on Image Processing, Chicago, 1998. 127139. 33 Kim H, Kang W, Shin J, Park S. Face Detectio
46、n Using Template Matching and Ellipse Fitting. IEICE Trans. Inf. & Syst., Vol.E38-D, No.11, pp20082011, Nov 2000. 34 Wang J, Tan T. A new face detection method based on shape information. PRL, vol. 21, pp. 463471, 00. 35 Frba B, Kblbeck C. Orientation Template Matching for Face Localization in Compl
47、ex Visual Scenes. International Conference on Image Processing ICIP2000, 251254, 2000. 36 Frba B, Kblbeck C. Real-Time Face Detection using Edge-Orientation Matching. Audio- and Video-based Biometric Person Authentication (AVBPA2001), 7883, 2001. 37 Frba B, Zink W. On the Combination of Different Te
48、mplate Matching Strategies for Fast Face Detection. Multiple Classifier Systems (MCS2001), 418428, 2001. 38 Govindaraju V, Srihari SN, Sher DB. A computational model for face location. In: Proc IEEE Conference on Computer Vision, Osaka, Japan, 1990. 718721. 39 Miao J, Yin BC, Wang KQ et.al. A hierarchical multiscale and multiangle system for human face detection in a complex back-ground using gravity-center template. Pattern Recognition, 1999, 32(10): 12371248. 40