基于不相关局部敏感鉴别分析的新生儿疼痛表情识别

针对新生儿疼痛表情识别中的特征提取问题,提出一种不相关局部敏感鉴别分析（Uncorrelated Locality Sensitive Discriminant Analysis,ULSDA）算法.首先,在局部敏感鉴别分析（LSDA）算法的基础上,通过附加投影向量正交性的约束条件,定义了ULSDA的目标函数;然后,推导出求解特征子空间正交投影向量的迭代公式;最后,将输入的高维图像数据投影到这个特征子空间,求出特征向量.ULSDA算法不仅继承了LSDA算法所具有的有监督、局部流形结构保持等特性,而且消除了LSDA算法所提取出的鉴别特征的相关性,从而增强了特征的鉴别能力,比LSDA算法具有更好的分类识别能力.在新生儿表情图像库上的疼痛表情识别实验结果表明,提出的ULSDA方法是有效可行的.当每类表情的训练样本图像为150幅时,采用ULSDA算法获得的平均识别率达到了82.07％,优于主成分分析（PCA）、线性鉴别分析（LDA）、局部敏感鉴别分析（LSDA）等特征提取方法.

Neonatal Pain Expression Recognition Based on Uncorrelated Locality Sensitive Discriminant Analysis

A novel feature extraction method called uncorrelated locality sensitive discriminant analysis （ULSDA） is proposed for neonatal pain expression recognition.First,on the base of locality sensitive discriminant analysis （LSDA）,the ULSDA objective function was defined by adding an orthogonal projection vectors based constraint to the LSDA objective function.Then,the iterative formulae to solve for orthogonal projection vectors were derived.At last,the original image data was projected onto the feature subspace for feature extraction.Thus,ULSDA shares the same supervised and local manifold structure preserving character as LSDA,but it outperforms LSDA in terms of eliminating the correlation between discriminant features,and as a consequence it is more effective than the LSDA for classification tasks.The experimental results show that the proposed ULSDA is effective and feasible for neonatal pain expression recognition.The average recognition rate achieved by the ULSDA algorithm is 82.07％ on the neonatal expression database with l =150,and is higher than those obtained by some popular methods such as principal component analysis （PCA）,linear discriminant analysis （LDA）,LSDA.