基于协同表示的图嵌入鉴别分析在人脸识别中的应用

Collaborative Representation Based Graph Embedding Discriminant Analysis for Face Recognition

鉴于协同表示投影方法中只考虑通过样本间协同表示关系构造邻接图,在投影后可能导致来自不同类别的许多样本聚集在一起影响识别结果,提出一种基于区分竞争和协同表示投影方法.首先,通过数据集中所有样本竞争协同地表示每个样本,计算样本间的相似性,以构造类内图描述类内样本的紧致性,同时构造惩罚图刻画不同类样本间的可分性;在此基础上,引入标签传播算法计算出数据集中无标签样本的软标签信息,以消除无标签样本对识别结果的影响;此外,使用非线性映射替换图嵌入框架中的线性内积,以解决原始数据集中的样本在低维空间中的线性不可分问题.在ORL,AR,FERET和Yale B人脸数据集上的实验结果表明,与CRP方法相比,所提方法最大识别率提升了1%~4%,尤其在噪声和模糊干扰下提高了2%~6%.

Considering the collaborative representation projection constructs the adjacent graph based on the collaborative representation relationship of samples and many samples from different classes may be gathered together after projection,a discriminative competitive and collaborative representation projection is proposed.Firstly,each sample is competitively and collaboratively represented by all samples in the dataset to calculate the similarity of samples.Then,an intraclass graph is constructed to characterize the intraclass compactness,and an interclass graph is built to characterize the interclass separability.On this basis,the label propagation algorithm is applied to calculate the soft label information of unlabeled samples to eliminate the influence of unlabeled samples on the recognition results.Additionally,the nonlinear mapping is used to replace the linear inner in the graph embedding framework to solve the linearly inseparable problem of the original samples in the low-dimensional space.Experimental results on ORL,AR,FERET and Yale B face datasets show that compared to the CRP method,the proposed method improves the maximum recognition rate by 1%~4%and improves 2%~6%in noise and blur results,respectively.