《基于深度学习的人脸识别技术综述.docx》由会员分享,可在线阅读,更多相关《基于深度学习的人脸识别技术综述.docx(17页珍藏版)》请在第壹文秘上搜索。
1、基于深度学习的人脸识别技术综述简介:人脸识别是计算机视觉研究领域的一个热点,同时人脸识别的研究领域非常广泛。因此,本技术综述限定于:一,在LFW数据集上(LabeledFacesintheWild)兼得优秀结果的方法二,是采用深度学习的方法前言1.FW数据集(LabeledFaCeSlntheWild)是目前用得最多的人脸图像数据库。该数据库共13,233幅图像,其中5749个人,其中1680人有两幅及以上的图像,4069人只有幅图像C图像为250.250大小的JPEG格式。绝大多数为彩色图,少数为灰度图。该数据库采集的是自然条件下人脸图片,目的是提高自然条件下人脸识别的精度。该数据集有6中评
2、价标准:Unsupervised;二,Image-restrictedwithnooutsidedata;三,Unrestrictedwithnooutsidedata:N,Image-restrictedwithlabel-freeoutsidedata:五,Unrestrictedwithlabel-freeoutsidedata:六,Unrestrictedwithlabeledoutsidedata:目前,人工在该数据集上的准确率在0S4270992(L在该数据集的第六种评价标准下(无限制,可以使用外部标注的数据),许多方法已经赶上(超过)人工识别精度,比方face+QeeplD3,E
3、QJ ST-HSOd 9nsHuman PerformanceHunanr unneledHunan, cropped Hman, inverse askHuman,funneled110.9920Humanzcropped110.9753Human,inversemask】0.9427Table7:Meanclassificationaccuracy0andstandarderrorOfthemeanSee.l8.2B.3.4.3e.0.7.8.9falsepositiverateFaceNet 等。Figure7:ROCcurvesaveragedover10foldsofView2.图-
4、/发一:人类在LFW数据集上的识别精度表二:第六种标推下,局部模型的识别准确率(详情参见IfW结果)Unrestricted,LabeledOutrideDataResult%Simileclassifiers0.84720.0041AttributeandSimileclassifiers110.85540.35MultiplLE*ConM0.84450.0046Associate-Predicte0.90570.56TOnVVS-Pete230.9310士0.0135Tom-VS-PeteAttribute30.93300.0128combinedJointBayesian%0.92420
5、.0108KighYmLBP0.95170.0113DFD330.84020.0044TLJointBay3an0.96330.0108r2011b190.91300.30Face+400.9950上0.36DpFac-esmbe10.9735OoO25ConvNet-RBM420.9252土0.38POOF-gradhist440.93130.0040PooFXOGy0.92800.0047FRFCn50.9645O.25DeepID460.97450.0026JUSSianFace70.9520.0066DeepID2480.99150.13530.93330.0124D IhC 2D-a
6、ligncd CroP image-plane. (C) Triangle visibility w.r.t. to the fitted 3D-2D camera; black triangles arc lc、 isible. (f) The 67 Iiducial points induced by the 3D model that arc using to direct the piecc-wisc ainc warpping, (g) The Iinal Irontalized crop, (h) A neu view generated hy the 3D model (not used in IhiS paper).图2一1人脸对齐的流程2.3深度神经网络hpmr: Otfhnr4IhrIkffFacrarikflccied byer aid i* Iulh-Iumeiitd l*)tn. Cu4n lhtr* naput (w eac