摘要
为了从大字典视频中跟踪和识别人脸,提出了一种基于平均序列的稀疏表示分类端到端识别方法。首先,利用所有可用视频数据和属于同一个人的人脸跟踪帧进行联合优化;然后,将严格的时间约束添加到l1-最小化;最后,运用人脸跟踪中所有单个帧重建各个身份,利用稀疏重建完成人脸分类。在You Tube人脸数据集上的实验验证了本文方法的有效性,在You Tube名人数据集和本文搜集的电影预告片数据集上的实验结果表明,相比几种较为新颖的分类方法,该方法取得了更高的识别精度,并且在拒绝不明身份上的准确率比SVM高8%。
To track and recognize face from video with large dictionary, an end-to-end recognition method based on sparse representation classification with mean sequences is proposed. Firstly, all the available video data and face tracking frame belonging to the same person is used to joint optimization. Then, strict time constraints are added into l^1-minimization. Finally, all individual frames in the human face tracking are used to reconstruction each identity, and sparse reconstruction is used to finish face recogni- tion. The effectiveness of proposed method is verified by experiments on YouTube database. Experimental results on YouTube Ce- lebrity database and movie trailers dataset searched by self show that proposed method has higher recognition accuracy than several advanced methods. The accuracy rate of proposed method is higher than SVM with 8% on refusing to unidentified.
出处
《电视技术》
北大核心
2015年第1期127-132,共6页
Video Engineering
基金
国家自然科学基金项目(61103143)
河南省科技厅自然科学研究计划项目(132300410276)
平顶山学院青年科研基金项目(PDSU-QNJJ-2013010)
关键词
平均序列
稀疏表示分类
视频检索
人脸跟踪
人脸识别
mean sequences
sparse representation classification
video retrieval
face tracking
face recognition