| 摘要: |
| 如何有效利用视频中人脸之间的时空连续性信息来克服人脸分辨
率低、图像尺度变化大和姿态、光照变化以及遮挡等问题是视频人脸识别的关键所在。提出
了一种基于流形学习的视频人脸性别识别算法。 该算法不仅可以通过聚类融合学习来挖掘
视频内在的连续性信息，同时能发现人脸数据中内在非线性结构信息而获得低维本质的流形
结构。在UCSD/Honda和自采集数据库上与静态的算法比较结果表明，所提算法能够获得更好
的识别率。 |
| 关键词: 视频人脸性别识别 流形学习 聚类融合 保局投影 支持向量
机 |
| DOI： |
|
| 基金项目: |
|
| An improved manifold-based face gender recognition algorithm for video |
| ZHANG Dan |
| () |
| Abstract: |
| How to fully utilize both spatial and temporal information in video to overcome
the difficulties existing in the video-based face recognition, such as low resol
ution of face images in video, large variations of face scale, radical changes o
f illumination and pose as well as occasionally occlusion of different parts of
faces, has become the research focus. In this paper, a novel manifold-based fac
e gender recognition algorithm for video(VG-LPP) using clustering is proposed, w
hich can discover more special semantic information hidden in video face sequenc
e, simultaneously well utilize the intrinsic nonlinear structure information to
extract discriminative manifold features.Comparison of VG-LPP with other algorit
hms on UCSD/Honda and the author′s own video databases shows that the proposed
a
pproach can perform better for video-based face gender recognition. |
| Key words: video-based face gender recognition manifold cluster
ing locality preserving projection support vector machine |
