基于协同表示标签传播的半监督正交鉴别分析算法

Collaborative Representation Label Propagation based Semi-supervised orthogonal Discriminative Analysis

近年来由于其表示的有效性,基于表示的分类方法,例如稀疏表示分类器(SRC)或者协同表示分类器(CRC)被广泛的应用于各种各样的识别任务.但是,SRC或者CRC的性能严重依赖于每类带标签训练样本的个数,当每类带标签的训练样本不够充分,SRC或者CRC的识别性能将会显著地下降.为了解决这个问题,文章[34]把协同表示技术引入到半监督学习方法中,提出了一种基于协同表示的标签传播算法,有效的利用了大量未标记样本的信息来进行标签传播.受此启发,把这种标签传播算法与正交鉴别分析算法相结合,提出了一种基于协同表示标签传播的半监督正交鉴别分析算法,目的是为了学习一个更好的鉴别子空间.不同于传统的半监督降维方法,所提算法首先利用这种标签传播算法将部分有标签数据的标签信息传递给不带标签的数据,之后利用传播后获得的全体软标签信息构造离散度矩阵实现鉴别分析,其次对鉴别投影施加正交约束,采用一种更加有效快速的迹比优化算法进行鉴别分析.大量的实验结果验证了所提算法的有效性.尤其在只存有少量标签样本的情况下,算法仍能保持良好的分类性能.

Recently,representation-based classification methods,such as sparse representation-based classification (SRC)and collaborative representation-based classification (CRC)have been developed and shown potential due to its effectiveness in various recognition tasks.However,the performance of SRC or CRC depend on the number of labeled training samples for each class.When the labeled training samples per class are insufficient,the recognition performance of SRC or CRC will decrease significantly.In order to address this problem,literature[34] introduced the collaborative representation into semi-supervised learning and proposed a semi-supervised label propagation method based on collaborative representation,which can explore the information of unlabeled data effectively.Inspired by this,we propose a collaborative representation label propagation based semi-supervised orthogonal discriminant analysis algorithm (CR-SODA),which combines the collaborative representation label propagation with the orthogonal discriminant analysis.Different from the existing semi-supervised dimensionality reduction algorithms,our method propagates the label information from the labeled data to the unlabeled data through this label propagation algorithm,and then construct the scatter matrices by using the soft label to perform discriminant analysis.In addition,we impose the projection matrix is orthogonal such that the solution of proposed method can be effective obtained by using a trace ratio optimization algorithm.Extensive experiments on face dataset verify the effectiveness of our algorithm,especially when limited number of images were labeled,our method can still maintain its good classification performance.