Face Recognition-based Automatic Tagging Scheme for SNS
Outline Introduction Motivation Related Work Methodology Face Detection Face Matching Experimental Results Conclusion 01
Introduction Tag is Metadata related with Internet Contents Used by Most Web 2.0-oriented websites Effective Information Classification System Applications Tag Cloud, Location Tag, Photo Tag Collaboration Tagging, Auto Tagging Problem Abuse / Misuse of Tag (spam tag) Troublesome work for normal user Not interesting 02
Motivation Automatic Tagging Using algorithms of pattern recognition Prevent inappropriate, unnecessary tagging Exact & Convenient Facial Tagging Tag whose name on his/her face Novel & Fun (image-based) Applicable to various field - People Searching - Categorize pictures by relationship 03
Related Work Face Recognition Methods Principal Component Analysis (M. Turk & A. Pentland, 1994) Linear Discriminant Analysis (S. Balakrishname & A. Ganapathiraju, 1998) Neural Network (J. E. Meng & W. Shiqian, 2002) Gabor Wavelet (L.C.Jain & U.Halici, 1999) Support Vector Machines (E. Osuna, & R. Freund 1997) 04
Related Work Face Detection Methods Knowledge-base Method (G. Yang & T.S. Huang, 1994) - Geometric features, Histogram Feature-based method (R. Kjeldsen & J. Kender, 1996) - Facial elements, texture, skin color, outline, or combination of them Template Matching Method (I. Craw et al., 1992) - Standard facial templates (created manually) - AAM(active appearance model), ASM(active shape model) Appearance-based Method (M. Turk & A. Pentland, 1991) - Using classifiers that is learned features of face - PCA, SVM, HMM, NN, etc. 05
Related Work Social Network Service (SNS) Internet service supporting social network on the web Sharing personal daily life or interesting things with UCC A value-added killer-app in Web 2.0 Cyworld, Facebook, Myspace, Twitter, Me2day, Mixi, etc. - Including Blog, Café and Virtual world(second Life) 06
Methodology 이충연남자 cylee Research Object Detect faces in pictures on the picture management system of SNS Match the face to the member data and display it Tag his/her name on the picture 07
Methodology Face Detection Haar-like features - Calculate the difference of the sum of pixels of areas inside the rectangle - The values indicate certain characteristics of a particular area of the image - Each feature type can indicate the existence of certain characteristics in the image, such as edges or changes in texture. edge Dark area Light area line center 08
Methodology Face Detection Using Integral image for fast computation - Summed area tables: 2D Lookup table with the same size of the original image - Each element of the Integral Image contains the sum of all pixels located on the up-left region of the original image - It allows to compute sum of rectangular areas in the image, at any position or scale, using only 4 lookups: sum = α β γ + δ A α B β C γ D δ 09
Methodology Face Detection Classifier Learning using AdaBoost - A weak classifier is called repeatedly in a series of rounds t = 1,, T - For each call, a distribution of weights D t is updated that indicates the importance of examples in the data set for the classification. - On each round, the weights of each incorrectly classified example are increased, so that the new classifier focuses more on those examples. 09
Methodology Face Matching Feature Extraction - Make a face vector from face image - Extract its eigenvector and eigenvalue using PCA - Since the feature seems like a face, it is dubbed eigenface { } Recognition & Matching - Compare the feature vector of input image with the feature vectors of learned image set - The face has minimum euclidean distance is the closest one - So, it returns the index number and seek the member data from DB 10
Experimental Results System Overview Detected facial region Recognized Profile Trimmed facial region Input tag list Activate face detection 11
Experimental Results Detection time average detection time : 366.96 ms 700 600 500 400 300 200 100 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Total subjects : 30 12
Experimental Results Detection rate 100 average detection rate : 90.8% 80 60 40 20 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Picture : 30, Total subject : 95 12
Experimental Results Learning time 1000 800 600 400 200 0 average learning time : 498.75 ms 782 638 512 759 335 424 289 251 3 4 5 6 7 8 9 10 Learning time 12
Experimental Results Recognition time average reconition time : 396.93 700 600 500 400 300 200 100 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Total recognition subject 12
recognition rate Experimental Results Recognition rate 100 average recognition rate : 76% 80 60 40 20 0 1(2) 2(3) 3(2) 4(2) 5(2) 6(2) 7(4) 8(4) subject (number of trial) 12