兒童面部識別的縱向評估.pdf

《兒童面部識別的縱向評估.pdf》由會員分享，可在線閱讀，更多相關《兒童面部識別的縱向評估.pdf（20頁珍藏版）》請在三個皮匠報告上搜索。

1、Longitudinal Evaluation of Child Face RecognitionSurendra Singh(Clarkson University),Stephanie Schuckers(UNC Charlotte)1Challenge in child face recognition due to non-linear cranial growthDeep Neural Network(DNN)models for adults may not always be applicable to children.ObjectiveAnalyze DNN performa

2、nce on the YFA(young face aging)dataset(age up to 8 years).Study specific changes in face features,e.g.,nose,mouth and eyes.Identify unique physiological factors contributing to childrens facial development.Enhance accuracy and effectiveness of face recognition(FR)systems for children.ValueUnderstan

3、ding Challenges in Child Face Recognition.Benchmarking FR Performance Based on Growth in Children.Discovering Specific Changes in Facial Features Impacting Matching Performance.Problem2YFA Database(Young Face Aging)The Young Face Aging(YFA)DatabaseNumber of subjects at each age in the YFA Database.N

4、umber of images at each age in the YFA Database.Contains face images of children aged 3-18 years.330 subjects,with an average of six collections per subject over eight years.Images of the same subject were collected every six months for eight years.The first collection image was used for enrollment

5、and verified against each subsequent collection over the eight-year period.The database includes 60 subjects with a total of 1,322 samples collected over eight years.Collected in a controlled environment with consistent indoor lighting,neutral expressions,and minimized pose variations.Manual annotat

6、ion to exclude extremely blurry images and challenging poses.3Prior work DatabaseLongest time gap Time intervalAccuracyModelECLF113 years6 monthsTAR at 0.1%FARFaceNet:84.55 PFE:98.90 ArcFace:99.38 COTS:99.62ITWCC-D112TAR at 0.1%FARFR Model:COTS FR-A:0.676 FR-B:0.598 FR-C:0.463 FR-D:0.434 FR-E:0.759

7、FR-F:0.738 FR-G:0.718 FR-H:0.695NITL13 2 years1 yearTAR at 0.1%FARCOTS:60.94 CLF143 years3 year TAR at 0.1%FARCOTS:49.33FaceNet:59.80CMBD15Rank-1 AccuracyPCA:38.8 LBP:28.8 LDA:71.3 Fine-tuned VGG-Face:83.0 Triplet CNN:72.7 Proposed CNN:85.1YFA8 years6 monthsTAR at 0.1%FARMagFace:95.484Prior work on

8、YFA database Face Recognition In Children:A Longitudinal Study 1TAR 0.1%FARModelT=6M T=12MT=18MT=24M T=30MT=36MFacenet-V1 95.894.8 92.584.382.776.0ArcFace 87.688.185.3 84.886.381.1ArcFace-Focal97.698.3 95.492.793.191.6 MagFace 98.298.3 98.097.297.394.9 In prior work,we evaluated multiple open-source

9、 DNN-based face recognition models and found that MagFace performed the best.Therefore,we selected MagFace for our further analysis.MagFace2:Training database:MS1M-V2 3 (5.8M images,85k identities)Evaluation database:LFW 4,CFPFP 5,AgeDB-30 6,CALFW 7,CPLFW 8,IJBB 9 and IJB-C 105Experimental Setup and

10、 Overall ResultsUse of MTCNN for accurate face detection and alignment.Feature extraction using MagFace 12.Input image size 112x112.ModelTAR 0.1%FARThresholdTAR 0.01%FARThresholdMagFace95.480.4582.250.56GenderTARThresholdMale94.870.45Female 95.800.45 Gender-Based TAR Performance 0.1%FAR6Methodology

11、TAR%at 0.1%FARCategorization of images into enrollment and verification samples based on age increments of 6 monthsEnrollment Stage:Any image in the dataset can be an enrollment image,categorized by age brackets.Verification Stage:Subsequent collections used for verification,evaluating performance o

12、ver time.7Analysis for age groups8Investigation of age-related trends in face recognition performance.Analyzing the impact of enrollment age on True Acceptance Rate(TAR)over time.Recognition accuracy declines significantly beyond a 4-year age gap.The(3-5)and(7.5-9)age groups experience the largest d

13、rop in TAR over time.Bootstrapping for performance evaluation 9Bootstrap Statistical technique for estimating sampling distributions.Resamples data to approximate uncertainty without assuming distribution.Provides robust estimates of parameters(e.g.,confidence intervals).To analyze TAR variability a

14、cross different age groups and verification periods.Utilized bootstrap resampling for robust estimation.Significant TAR differences observed across age groupsFacial Feature Growth AnalysisMeasured growth in facial distances(e.g.,nose length,chin size,mouth width).Normalized distances using inter-eye

15、 distance.Significant growth observed in features between ages 4-16,impacting recognition accuracy.10Face Features-Relative nose length to distance between the eyes As children age,the relative distance from the nose to the eyes increases.We are continuing to analyze other facial features,such as th

16、e mouth and chin.Distance calculation is by median.Distance equation:Distance between feature points(in pixel)Distance between middle point of eyes 11Face Cropping&Recognition AccuracyEvaluated multiple face cropping algorithms:MTCNN,RetinaFace,OpenCV,DLIB,etc.MTCNN and RetinaFace provided the highe

17、st recognition accuracy for children.OpenCV showed strong performance for adults but lower accuracy for children.Face cropping with difference face cropping a:Dlib,b:Mediapipe,C:MTCNN,D:OpenCV,e:Retinaface,f:SSD12Conclusions and Future work Key Findings:Average TAR:(0.5-2)years:98.52%(2.5-4)years:95

18、.68%Significant decline post 4-year age difference.Notable TAR declines observed in younger age groups(e.g.,3-5 years:63.1%)Conclusion:Accuracy fluctuations highlight the complexity of age-related biometric performance and the need for adaptive recognition models.Study limitations include uneven age

19、 distribution,lighting inconsistencies,demographic biases,and reliance on a single face-matching algorithm(MagFace).Findings contribute to understanding child facial recognition and its implications for applications like missing child identification.Future WorkExpand dataset diversity to improve gen

20、eralizability across age groups and ethnic backgrounds.Investigate deep learning-based models to adaptively account for age progression in child biometrics.Assess performance of different face-matching algorithms and explore fine-tuning approaches.Conduct real-world evaluations to address operationa

21、l challenges in uncontrolled environments.13Reference 1 Bahmani,K.,&Schuckers,S.(2022,April).Face recognition in children:a longitudinal study.In 2022 International Workshop on Biometrics and Forensics(IWBF)(pp.1-6).IEEE.2 Meng,Q.,Zhao,S.,Huang,Z.,&Zhou,F.(2021).Magface:A universal representation fo

22、r face recognition and quality assessment.In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition(pp.14225-14234).3 Jiankang Deng,Jia Guo,Niannan Xue,and Stefanos Zafeiriou.ArcFace:Additive angular margin loss for deep face recognition.In IEEE Conference on Computer Visi

23、on and Pattern Recognition,pages 46904699,2019.1,2,3,4,5,6,11,12 4 Gary B Huang,Marwan Mattar,Tamara Berg,and Eric Learned-Miller.Labeled faces in the wild:A database for studying face recognition in unconstrained environments.Technical Report 07-49,University of Massachusetts,Amherst,20075 Soumyadi

24、p Sengupta,Jun-Cheng Chen,Carlos Castillo,Vishal M Patel,Rama Chellappa,and David W Jacobs.Frontal to profile face verification in the wild.In IEEE Winter Conference on Applications of Computer Vision(WACV),pages 19.IEEE,20166 Stylianos Moschoglou,Athanasios Papaioannou,Christos Sagonas,Jiankang Den

25、g,Irene Kotsia,and Stefanos Zafeiriou.AgeDB:the first manually collected,in-the-wild age database.In IEEE Conference on Computer Vision and Pattern Recognition Workshop,pages 5159,2017.57 Tianyue Zheng,Weihong Deng,and Jiani Hu.Crossage LFW:A database for studying cross-age face recognition in uncon

26、strained environments.arXiv preprint arXiv:1708.08197,2017.8 Tianyue Zheng and Weihong Deng.Cross-pose LFW:A database for studying cross-pose face recognition in unconstrained environments.Beijing University of Posts and Telecommunications,Tech.Rep,5,20189 Cameron Whitelam,Emma Taborsky,Austin Blant

27、on,Brianna Maze,Jocelyn Adams,Tim Miller,Nathan Kalka,Anil K Jain,James A Duncan,Kristen Allen,et al.IARPA Janus benchmark-B face dataset.In IEEE Conference on Computer Vision and Pattern Recognition Workshop,pages 9098,201710 Brianna Maze,Jocelyn Adams,James A Duncan,Nathan Kalka,Tim Miller,Charles

28、 Otto,Anil K Jain,W Tyler Niggel,Janet Anderson,Jordan Cheney,et al.IARPA Janus benchmark-C:Face dataset and protocol.In International Conference on Biometrics,pages 158165.IEEE,2018.14Reference 11 P.K.Chandaliya,Z.Akhtar,and N.Nain,“Longitudinal analysis of mask and no-mask on child face recognitio

29、n,”in Proceedings of the Thirteenth Indian Conference on Computer Vision,Graphics and Image Processing,2022,pp.17.12 N.Srinivas,K.Ricanek,D.Michalski,D.S.Bolme,and M.King,“Face recognition algorithm bias:Performance differences on images of children and adults,”in Proceedings of the IEEE/CVF confere

30、nce on computer vision and pattern recognition workshops,2019,pp.00.13 L.Best-Rowden,Y.Hoole,and A.Jain,“Automatic face recognition of newborns,infants,and toddlers:A longitudinal evaluation,”in 2016 International Conference of the Biometrics Special Interest Group(BIOSIG).IEEE,2016,pp.18.14 D.Deb,N

31、.Nain,and A.K.Jain,“Longitudinal study of child face recognition,”in 2018 International Conference on Biometrics(ICB).IEEE,2018,pp.225232.15 S.Siddiqui,M.Vatsa,and R.Singh,“Face recognition for newborns,toddlers,and pre-school children:a deep learning approach,”in 2018 24th International Conference

32、on Pattern Recognition(ICPR).IEEE,2018,pp.31563161.16 K.Bahmani and S.Schuckers.Face recognition in children:A longitudinal study.In 2022 International Workshop on Biometrics and Forensics(IWBF),pages 16.IEEE,2022.15Backup Slide 16Face features 17Face featuresLongitudinal Growth Measurements of Faci

33、al Features Between Ages 4 and 6,Highlighting Changes in Nose Length,Mouth Dimensions,Chin Size,and Horizontal Mouth Length.Male:Longitudinal Growth Measurements of Facial FeaturesFemale:Longitudinal Growth Measurements of Facial Features18Experimental Setup and Overall ResultsUse of MTCNN for accur

34、ate face detection and alignment.Cropped faces resized to 224x224 pixels for consistency in analysis.Feature extraction using MagFace 12.Input image size 112x112.ModelTAR 0.1%FARThresholdTAR 0.01%FARThresholdMagFace95.480.4582.250.56GenderTARThresholdMale94.870.45Female 95.800.45Age GapTARSubjects2 years98.52%3234 years95.68%1996 years87.24%1468 years71.32%126 Age-Based TAR Performance 0.1%FAR Gender-Based TAR Performance 0.1%FAR19DLIBMediapipeMTCNNOpenCVRetinafaceSSD20