Font Size: a A A

Face Analysis And Verification Under Uncontrolled Conditions

Posted on:2015-03-01Degree:DoctorType:Dissertation
Country:ChinaCandidate:F Y SongFull Text:PDF
GTID:1108330479475940Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Among several popular biometrics including fingerprint, speech, and eye iris recognition, Face Recognition Technology (FRT) shows one prominent feature that it does not require aid (or consent) from the test subjects. With this advantage, FRT uniquely performs well in mass identification, i.e., identify individuals among the crowd, and has potential applications in public security. On the other hand, research on FRT contributes to both academic and application developments in image understand-ing related subjects, and therefore has drawn extensive attention from the domains of Computer Vision, Pattern Recognition, Machine Learning, and also Statistics in recent several decades. Numerous meth-ods have been proposed, some of which perform nearly perfect under constrained conditions. However, general face recognition under uncontrolled conditions still keeps unsolved due to complex appearance caused by great variations in lighting, expression, pose, and also partial occlusion. The focus of this paper is just towards robust face understanding under uncontrolled conditions.This paper focuses on two key topics involved in uncontrolled face analysis and verification, i.e., facial landmarks localization and face representation learning. To address the first topic, accurate and robust eye localization under uncontrolled conditions is investigated, together with eye states analysis as its subsequent task. Focusing on the second topic, two typical types of face representation are dis-cussed, including supervised attribute representation learning, and clustering representation learning as an unsupervised way. Finally, focusing on the final task, i.e., robust and accurate face analysis and ver-ification, a novel face verification model that owns adaptability to testing face pairs, is proposed. The main contributions can be summarized as follows.(1) A comprehensive survey on typical approaches for eye localization under unconstrained con-ditions is given and organized in three main perspectives. On the basis of this, a robust and efficient eye localization framework is proposed, which involves several key issues including preprocessing and postprocessing. Eye localization is also discussed from the view of general object recognition, expect-ing ideas of general object detection may contribute new ways of accurate and robust eye localization. Finally, performance comparison and discussion of popular approaches for eye localization is given, and also the problems existed in performance evaluation.(2) A general appearance-based framework for judging eye states (open/closed) is given. His-togram of Principal Oriented Gradients is proposed to address the unstable computation of pixel gradi-ents caused by image noises, and is further extended to address scales variation. Further, feature fusion strategy is adopted to achieve robust appearance description under uncontrolled conditions. Promising results are verified on eye states recognition in real scenarios.(3) Under the framework of supervised representation learning, attribute relationship is exploited for improved face verification. Since it is difficult to measure attribute relationship directly, a new distributed representation in the subject space for each attribute is built by measuring its correlation with each subject. Then the attribute-graph can be calculated in this new representation space. Motivated by the insight that the more related attributes should contribute similar to final recognition task, attribute-graph is further formulated into Laplacian matrix as a regularization term to constrains the size of hypothesis space in large margin classification framework. The proposed approach is further extended to work on general continuous attribute data (i.e., general features). Extensive experiments verified the effectiveness of the proposed methods.(4) Under the framework of unsupervised representation learning, clustering representation is dis-cussed. A discriminant and distributed representation space is built by exploiting large amount of Web faces through clustering technique with low costs. Further, addressing limited identity information in testing face pair since their subjects are never seen before, a test-instance adaptive model is learned for each testing face pair, and then turns the face pair verification problem into verifying the new trained test-instance adaptive model by checking its prediction behavior on a held-out set of face pairs. In this way, face verification on model level exploits much information that beyond a single face pair.
Keywords/Search Tags:Face Verification, Uncontrolled Condition, Eye Localization, Supervised Represen- tation Learning, Attribute Learning, Attribute Relationship, Unsupervised Representation Learning, Clustering Representation, Appearance Model, Structural Model, One Shot
PDF Full Text Request
Related items