中文摘要：

为解决鉴别稀疏邻域保持嵌入（DSNPE）算法中类间离散度构造复杂的问题,提出了一个新的维数约简算法即鉴别稀疏局部保持投影的人脸识别算法（DSLPP）。首先利用样本集中各类样本的平均向量构造字典,通过保持各类样本平均向量的稀疏重构关系,提出一个新的无参数类间离散度;再通过同时最大化类间离散度和同时最小化类内紧凑度的准则来寻找最优投影方向;最后采用最近邻分类器进行人脸分类识别。由于所采用的类间离散度最大限度地扩大了不同类别中样本之间的差异,因此DSLPP算法具有更强的类间判别力,其识别率得到了明显提高;此外,字典的简化构造降低了算法的计算复杂度。在Yale、UMIST和AR人脸库上的实验结果表明：DSLPP算法在Yale、UMIST库上的平均识别率及AR库上的最高识别率分别达83.38%、95.72%和83.71%,较其他传统方法的识别率有明显提高;在UMIST库上的实验结果表明,DSLPP算法较DSNPE算法的平均计算时间减少了81.7%。

英文摘要：

A new face recognition algorithm,i.e.a discriminant sparse locality and preserving projection algorithm（DSLPP）,is proposed to solve the problem that the construction betweenclass scatters is too complex in the discriminant sparse neighborhood and preserving embedding（DSNPE）method.A novel between-class scatter is constructed by using the mean vector of each class as dictionary and preserving the sparse reconstructive relationship of mean face.Then,an optimal projection matrix is obtained by maximizing the between-class scatter and minimizing the with-class compactness simultaneously.The nearest neighbor classifier is finally used for face recognition.The proposed between-class scatter maximizes the difference of samples between different classes and has more discriminant power,so that the recognition rate of the proposed algorithm is markedly improved.Moreover,the computational complex of the DSLPP algorithm is reduced because of the simple design of the dictionary.Experimental results show that theDSLPP algorithm achieves average recognition rates 83.38% and 95.72% on Yale,and UMIST face database respectively,and a maximal recognition rate 83.71% on AR face database,and that the recognition rates are obviously higher than the recognition rates of some conventional methods.The experimental results on UMIST face databases also show that the average computation time of the DSLPP algorithm is less 81.7%than that of the DSNPE algorithm.