ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2014, Vol. 51 ›› Issue (12): 2663-2670.doi: 10.7544/issn1000-1239.2014.20131200

• 人工智能 • 上一篇    下一篇



  1. (燕山大学信息科学与工程学院 河北秦皇岛 066004) (
  • 出版日期: 2014-12-01
  • 基金资助: 

Low-Rank Relaxed Collaborative Representation Combined with Global and Local Features for Face Recognition

Zhang Pan, Lian Qiusheng   

  1. (School of Information Science and Engineering, Yanshan University, Qinhuangdao, Hebei 066004)
  • Online: 2014-12-01

摘要: 目前的人脸识别算法经常忽视训练过程中噪声的影响,训练数据受到污染时识别性能会明显下降.针对该问题,提出了融合整体与局部特征的低秩松弛协作表示的人脸识别算法.通过低秩分解抑制训练样本的稀疏噪声,得到更加有效的人脸信息.利用松弛协作表示得到判别性更强的编码系数,增强人脸识别系统的判别性.为进一步提高识别率,提取局部特征的同时引入整体特征,运用整体特征和局部特征共同表示人脸图像.实验结果表明,尽管训练过程、测试过程都受到噪声污染,提出的算法对有光照、遮挡及表情变化的正面人脸图像的识别具有很好的鲁棒性,比现有的识别算法拥有更高的识别率.

关键词: 人脸识别, 低秩分解, 松弛协作表示, 整体特征, 局部特征

Abstract: Face recognition is among the most popular biometric approaches due to its low intrusiveness and high uniqueness. However, recent face recognition algorithms such as sparse representation-based classification algorithms, often ignore the noises during training process. When training samples are corrupted, recognition performance will significantly degenerate. Therefore, face recognition is an active yet challenging topic in computer vision application. To solve this problem, this paper proposes a novel low-rank relaxed collaborative representation algorithm which is combined with global and local features. The low-rank decomposition algorithm separates sparse noise from training samples to get more effective face information. Suppressing the effect of sparse noise is essential for face recognition system. Meanwhile, considering the fact that the different features in a sample should contribute differently to the pattern classification, this paper uses relaxed collaboration representation to get more discriminating coding coefficients. The discrimination of face recognition system will be enhanced. In addition, combining local features with global features can further improve the recognition rate. The proposed algorithm contains all of the above advantages. Experimental results on face databases show that, although both training and testing image data might be corrupted due to occlusion and disguise, the proposed algorithm is very competitive with state-of-the-art recognition approaches on effectiveness and robustness.

Key words: face recognition, low rank decomposition, relaxed collaborative representation, global features, local features