一种光滑局部敏感鉴别分析方法

Smooth Locality Sensitive Discriminant Analysis Method

传统的局部敏感鉴别分析方法未考虑原有图像样本像素关系信息,识别效果受到影响。为此,提出一种光滑局部敏感鉴别分析方法。针对图像样本构造一个基于离散拉谱拉斯图的正则化项,该正则化项包含图像像素关系的先验信息,并将其嵌入到局部敏感鉴别分析的目标函数中,使抽取的特征具有空间光滑的特性,从而增强局部敏感鉴别分析算法的泛化能力。在ORL和IMDB人脸数据集上的实验结果证明了该方法的有效性。

Locality Sensitive Discriminant Analysis（LSDA） is a recent proposed supervised feature extraction algorithm.LSDA can not only utilize the class information but also consider the the intrinsic geometrical structure of the data.However,LSDA is a vector based method so it neglecteds the spatial correlation of the pixels in the image,and its performance may be degraded in this case.To solve the problem,this paper proposes a Smooth LSDA（S-LSDA） method.It introduces a spatially smooth regularization which incorporates the spatial correlation information into the objective function of LSDA.It shows that the derived coefficients are spatially smooth and the extracted features are more effective for classification.Experimental results on face image databases show the effectiveness of the proposed algorithm