基于二维局部保持鉴别分析的特征提取算法

发布时间：2018-07-07 12:41

本文选题：人脸识别 + 表情识别　；参考：《南京邮电大学学报(自然科学版)》2014年05期

【摘要】：提出了一种二维局部保持鉴别分析(Two-dimensional Locality Preserving Discriminant Analysis,2D-LPDA)特征提取算法。该算法直接对图像矩阵进行运算而不需要将矩阵转化为向量后进行运算,较好地保持了图像相邻像素之间的空间结构关系;在LPP算法的基础上,利用训练样本的类别信息计算二维类间散度矩阵和二维类内散度矩阵,并在2D-LPDA的目标函数中引入最大间距准则(Maximum Margin Criterion,MMC),从而求得具有良好鉴别能力的投影向量,同时还避免了小样本情况下矩阵的奇异性问题。通过在ORL人脸图像库上的人脸识别和新生儿面部图像库上的疼痛表情识别实验,验证了所提出的算法的有效性。
[Abstract]:A feature extraction algorithm based on Two-dimensional locality preserving discriminant Analysis (2D-LPDA) is proposed. The algorithm directly computes the image matrix but does not need to transform the matrix into a vector, so the spatial structure relationship between the adjacent pixels of the image can be better maintained. On the basis of the LPP algorithm, The class information of the training sample is used to calculate the two-dimensional inter-class divergence matrix and the two-dimensional intra-class divergence matrix, and the maximum margin criteria (MMC) is introduced into the objective function of 2D-LPDA to obtain the projection vector with good discriminant ability. At the same time, the singularity of matrix is avoided in the case of small sample. The effectiveness of the proposed algorithm is verified by the experiments of face recognition on ORL face image database and pain expression recognition on newborn facial image database.
【作者单位】：南京邮电大学通信与信息工程学院;台湾大学电机资讯学院;东南大学水声信号处理教育部重点实验室;
【基金】：国家自然科学基金(61071167) 江苏省高校优秀中青年教师和校长境外研修计划资助项目
【分类号】：TN911.73;TP391.41

【参考文献】