摘要
基于稀疏表示的人脸识别问题希望字典同时具有良好的表示能力和较强的辨识性。采用判别式K-SVD(D-ksvd)算法,可训练得到较好的字典和线性判别函数,但该算法中的初始化字典是从各类样本中选择部分样本经K-SVD方法得到的,不能较完整地表示所有样本的特性,影响了基于该初始字典的训练字典的表示能力和分类器的辨识性。在字典初始化方法上进行了改进,先训练类内字典再级联成新的初始化字典,由于类内训练字典是各类别的优化字典,降低了训练字典的误差,提高了训练字典与线性分类器的判别性,在保持较快识别速度的同时,提高了人脸识别率。
Face recognition problem based on sparse representation attempts to obtain a dictionarywith both good represent power and effective discriminative ability. The discriminative K-SVD algorithm (D-ksvd) based on sparse representation is a dictionary training method which satisfies the above re quirement jointly. However, the initialized dictionary of the D-ksvd algorithm is trained from some sam- ple selected from the training data using K-SVD, which cannot represent the training data completely, and increases the residual of the initialization dictionary. The face recognition rate will be affected by the aforementioned problem. The algorithm proposed in this paper improves the initialization method of D ksvd algorithm. The dictionaries are trained in every category and join together to form a new initialized dictionary. Every learned dictionary is the optimized dictionary in each category, which decreases the re- sidual of the trained dictionary, and increases the discriminative ability of the trained dictionary and the linear classifier. The face recognition rate is increased and the average recognition speed is fast.
出处
《计算机工程与科学》
CSCD
北大核心
2014年第1期150-154,共5页
Computer Engineering & Science
基金
国家自然科学基金资助项目(60875016)
关键词
人脸识别
改进D-ksvd
稀疏表示
训练字典
face recognition
improved D-ksvd
sparse representation
dictionary training