并行子空间学习方法及其大规模图像识别应用研究

Subspace learning is an effective image feature extraction and recognition technique. However, in its real-world applications, how to recognize the large-scale image sample set is a difficult issue. In order to reduce the computational time and improve the recognition performance of subspace learning technique under the situation of large-scale image recognition, we propose the parallel subspace learning approaches: (1) First, we develop a parallel subspace learning framework, which divides the sample set into several subsets by designing two random data division strategies that are Equal Data Division (EDD) and Unequal Data Division (UDD). These two strategies correspond to equal and unequal computational abilities of nodes under parallel computing environment. The graph embedding method is employed to provide a general formulation for the developed framework; (2) Under this framework, we propose a supervised parallel linear discriminant analysis approach, which computes linear discriminant features of each subset and selects features with the largest Fisher scores for classification; (3) Under this framework, we propose an unsupervised parallel locality preserving projection approach, which computes locality preserving projection features of each subset and selects features with the smallest Laplacian scores for classification. Face recognition and palmprint recognition are two major application fields of subspace learning technique. We verify the proposed approaches by constructing large-scale face and palmprint datasets.

子空间学习是一种有效的图像特征提取和识别技术。在该技术的实际应用中，如何识别大规模的图像样本集是一个难点。为了降低大规模图像识别中子空间学习方法的计算代价、提高识别效果，本项目提出了并行子空间学习新方法：(1)首先构造并行子空间学习框架，即把原始样本集随机划分成多个子集，根据并行计算环境中各节点的计算能力相等或者不等的情况，设计了等分和不等分数据划分策略，然后使用图嵌入表示方法给出学习框架；(2)根据框架，提出有监督的并行线性鉴别分析方法，即分别计算每个子集的线性鉴别特征，选择Fisher鉴别值大的特征用于分类。(3)根据框架，提出无监督的并行局部保留映射方法，即分别计算每个子集的局部保留映射特征，选择Laplacian值小的特征用于分类。人脸识别和掌纹识别是子空间学习技术应用比较多的领域。本项目将通过构造大规模的人脸图像库、掌纹图像库来验证所提出的方法。

子空间学习是一类重要的图像特征提取和识别技术。本项目实现了并行子空间学习方法，深入开展了图像特征提取技术的研究，提出了多种新方法，并应用到人脸识别等生物特征识别领域，取得了良好的实验效果。在本项目资助下，发表SCI检索论文8篇，EI检索论文10篇，包括国际权威会议CVPR、AAAI、IJCAI等和国际权威期刊IEEE Transactions on Image Processing、Pattern Recognition等；培养博士、硕士研究生10余名；参加了多个学术会议、进行了广泛的学术交流。