图像空间中的鉴别型局部线性嵌入方法

Discriminative locally linear embedding in image space

为了更好地利用图像的空间关系和类信息来提高局部线性嵌入的性能,提出一种针对图像识别的鉴别型局部线性嵌入算法,并应用于人脸识别。首先,利用自适应图像欧氏距离构建近邻矩阵,计算得到的权重矩阵,再由权重矩阵重构特征,然后重构出数据内在的低维空间,最后利用线性判别分析引入类信息解决局部线性嵌入算法对测试样本无法重构以及分类的缺陷。实验基于FRAV2D和ORL人脸数据库,分析了图像欧氏距离和自适应图像欧氏距离算法提取图像空间信息的能力,并将本文提出的算法与目前已经广泛使用的人脸识别算法进行比较,其结果表明了鉴别型局部线性嵌入算法能更好地保留图像流形结构和类信息,显著提高人脸识别准确率。

In this paper, a discriminative locally linear embedding algorithm on image recognition, which considers spatial relationship of pixels and class information in order to improve the performance of locally linear embedding （LLE） , is presented. First, neighbor matrix, which is used to compute weight matrix, is constructed by adaptive image Euclidean distance, and features are reconstructed using the weight matrix. And then intrinsic lower-dimensional space of data is reconstructed. Finally, linear discriminant analysis is utilized to introduce class information to solve the defects that LLE can＇t reconstruct test samples and classify. Experiments are carried on FRAV2D and ORL databases. Comparing our proposed algorithm with popular algorithms in face recognition, these results show that a discriminative LLE can keep the best manifold structure and class information, and improve the accuracy of face recognition.