Multiscale binarised statistical image features for symmetric face matching using multiple descriptor fusion based on class-specific LDA
详细信息 查看全文
- 作者：Shervin Rahimzadeh Arashloo
- 关键词：Unconstrained face recognition ; Face image representation ; Binarised statistical image features ; Multiscale representation
- 刊名：Pattern Analysis and Applications
- 出版年：2017
- 出版时间：February 2017
- 年：2017
- 卷：20
- 期：1
- 页码：113-126
- 全文大小：
- 刊物类别：Computer Science
- 刊物主题：Pattern Recognition;
- 出版者：Springer London
- ISSN：1433-755X
- 卷排序：20
文摘
Local binary image coding for face image representation is established as a successful methodology mostly popularized by the well-known local binary pattern operator (LBP) and its variants. In this paper, an alternative learning-based binary image coding scheme is introduced which operates by projecting local image patches linearly onto a subspace using learnt filters. Most importantly, independent binarisation of filter responses is justified theoretically using independent component analysis in the filter learning stage. The extension of the method to a multiscale framework makes the feature capable to capture image content at multiple resolutions, improving its expressive power. Taking a local feature-based approach, the coded images are summarised regionally by histograms exploiting dense correspondences between images. A discriminative face image descriptor is constructed next by projecting the regional multiscale histograms onto a class-specific LDA space. The proposed discriminative descriptor can be learnt in an unsupervised fashion and hence perfectly suited for face recognition in unconstrained settings, including the unseen face pair matching task. Finally, the proposed MBSIF descriptor is combined with two state-of-the-art face image representations, namely the multiscale LBP and local phase quantisation features to further enhance the accuracy. The proposed approach has been evaluated extensively on the extended Yale B, LFW, FERET and the XM2VTS databases in various scenarios and shown to perform very favourably compared to the state-of-the-art methods.