稀疏表示的特征选择方法在人脸性别和种族识别中的应用

APPLYING SPARSE REPRESENTATION-BASED FEATURE SELECTION TO FACE RECOGNITION FOR GENDER AND RACE

针对人脸性别和种族识别中出现的特征冗余问题,提出一种基于稀疏表示的特征选择方法 SRFS(Sparse Representation based Feature Selection),构建了以区分类别为核心的特征选择框架模型。该模型将样本的特征空间的线性组合映射到不同类标所标记的类别,利用截断牛顿迭代法求解模型的非负稀疏解,进而通过稀疏值从特征空间中选择类区分度大且信息丰富的特征子集。通过实验,与一些比较经典的特征选择方法 Fisher Score、mRMR进行比较,验证了该方法的有效性,能够选择有效的特征,不仅降低特征的维数,而且获得了较高的识别率。

Aiming at the problem of feature redundancy in face recognition for gender and race, we propose a sparse representation-based feature selection method （SRFS） and build a feature selection framework model which takes the category differentiation as the core. The model maps the linear combination of sample＇ s feature space onto those categories marked by different class labels, and utilises the truncated Newton iteration method to obtain the nonnegative sparse solution, and then selects the feature subset with big differentiation and rich information from feature space by sparse values. Through experiments the SRFS is compared with these rather typical feature selection methods such as Fisher Score and mRMR, and is verified its effectiveness. SRFS can select effective feature subset, apart from reducing the dimensions of feature, it also achieves higher recognition rate.