Research On Unconstrained Face Identification

Face recognition has gained more and more attention over the last few years, due to its many application prospects in various domains such as authentication, vedio surveillance and human-computer interaction. In the past several years, face recognition methods were focused on the problems of face verification in the unconstrained scenarios and face identification in controlled or semi-controlled environment rather than the face identification. However, with the development of mobile internet technology, demanding for the one-to-more face recognition based applications is increasing which is not limited in vedio surveillance and other traditional application fields. This paper studies the face identification methods in uncontrolled environment, and the research work is carried out in two aspects:feature extraction and classifier design.In the feature extraction stage, most approaches need to align the faces firstly and then extract the features. However, for the uncontrolled face images, it is difficult to do alignment for the face images with partial deletion, so an align-free face image descriptor named RootHOG is proposed in this paper. To develop the representation technique Multi-Keypoints Descriptor (MKD), this paper propose Multi-Keypoint Multi-Descriptor (MKMD) and Multi-scale Multi-Keypoint Multi-Descriptor (MSMKMD). The experimental results show that MKMD can not only solve the problem of aliasing effect but also retain as much the high frequency details as possible. The MSMKMD introduces multi-scale information to represent a face image and increases the number of descriptors of one class in gallery dictionary, which improve the identification rate by the construction of such a local descriptor-based over completed dictionary.For the classifier aspects, this paper propose to combine two strong classifiers SRC and SVMs to construct a two layer cascade classifier, the SRC takes the role of a rough classifier and SVMs takes the fine one. Besides, this paper use different types of feature descriptor in each classification step and fuse the texture based descriptors LBP, TPLB and FPLBP in the second round classification.To better study the effectiveness of the proposed methods in this paper, two public databases LFW and PubFig are used in the experimental process. Comparing with other state-of-the-arts on the two databases and taking the MKD-SRC as a baseline, the superiority of the proposed methods is obvious, for the recognition rate is advanced by almost15percent on both databases.