ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2018, Vol. 55 ›› Issue (8): 1760-1772.doi: 10.7544/issn1000-1239.2018.20180364

所属专题: 2018数据挖掘前沿进展专题

• 人工智能 • 上一篇    下一篇



  1. (南京师范大学计算机科学与技术学院 南京 210023) (
  • 出版日期: 2018-08-01
  • 基金资助: 
    国家自然科学基金重点项目(61432008);国家自然科学基金项目(61272222) This work was supported by the Key Program of the National Natural Science Foundation of China (61432008) and the National Natural Science Foundation of China (61272222).

Fast Low-Rank Shared Dictionary Learning with Sparsity Constraints on Face Recognition

Tian Ze, Yang Ming,Li Aishi   

  1. (College of Computer Science and Technology, Nanjing Normal University, Nanjing 210023)
  • Online: 2018-08-01

摘要: 字典学习是重要的特征表示方法之一,在人脸识别等方面有广泛的应用,特别适合解决姿态变化下的人脸识别问题,因而倍受研究者的关注.为有效增强字典的判别能力,研究者结合领域知识和抗噪等策略提出大量的字典学习模型,其中包括最近提出的同时进行降维和字典学习的方法,但这些方法侧重考虑样本中特定类的信息,未能有效考虑训练样本间的共享信息.因此,提出了一种稀疏约束下快速低秩共享的字典学习方法.该方法采用降维和字典联合进行学习的方式,并嵌入Fisher判别准则获得特定类字典和编码系数,同时施加低秩约束获得低秩共享字典,以此增强字典和编码系数的判别能力.此外,运用Cayley变换保护投影矩阵的正交性来获得紧凑的特征集合.在AR,Extended Yale B,CMU PIE和FERET四个数据集上的人脸识别实验验证所提方法的优越性.实验结果表明所提方法在表情变化下的人脸识别具有很强的鲁棒性,并对光照起到了抑制作用,尤其适合解决光照、表情变化下的小样本问题.

关键词: 人脸识别, 字典学习, 稀疏约束, 低秩模型, 共享特征

Abstract: Dictionary learning is one of the most important feature representation methods. It has a wide range of applications in face recognition and other aspects. It is particularly suitable for solving face recognition problems under the change of pose, and has attracted much attention from many researchers. In order to enhance the discriminative ability of dictionary, researchers have put forward a large number of dictionary learning models in combination with domain knowledge and anti-noise strategies, including the recently proposed methods for simultaneous dimensionality reduction and dictionary learning, but these methods focus on the specific-class samples and fail to consider the sharing information between training samples. Therefore, we propose a fast low-rank shared dictionary learning with sparsity constraints approach. The method learns dimensionality reduction and dictionary jointly, and embeds Fisher discriminant criteria to obtain specific-class dictionary and coding coefficients. At the same time, we enforce a low-rank constraint to obtain the low-rank shared dictionary to enhance the discriminative ability of dictionary and coding coefficients. In addition, the Cayley transform is used to protect the orthogonality of the projection matrix to catch a compact feature set. Face recognition experiments on AR, Extended Yale B, CMU PIE, and FERET datasets demonstrate the superiority of our approach. The experimental results show that the proposed method has strong robustness to face recognition under facial expression changes, and plays an inhibitory role in lighting. It is especially suitable for solving small sample problems under illumination and expression changes.

Key words: face recognition, dictionary learning, sparsity constraints, low-rank models, shared features