一种组合特征抽取的新方法

该文提出了一种基于特征级融合的特征抽取新方法，首先，给出了一种合理的特征融合策略，即利用复向量给出组合特征的表示，将特征空间从实向量空间拓广到复向量空间，然后，发展了具有统计不相关性的鉴别分析的理论，并将其用于复向量空间内最优鉴别特征的抽取，最后，在Concordia大学的CENPARMI手写体阿拉伯数字数据库以及南京理工大学NUST603HW手写汉字库上的试验结果表明，所提出的组合特征抽取方法不仅具有很强的维数压缩能力，而且较大幅度地提高了识别率。

A Novel Feature Extraction Method Based on Feature Integration

Feature level fusion plays an important role in the process of data fusion. The advantage of feature level fusion lies in two aspects: Firstly, it can derive the most discriminatory information from original multiple feature sets involved in fusion; Secondly, it enables to eliminate the redundant information within the original feature sets and to make it possible for the decision in real time. The classical feature fusion based feature extraction approach is to group two sets of feature vectors into one union vector (or supervector) and then based on them for feature extraction. This approach is always computationally expensive due to the high dimensional supervectors resulting from integration. To overcome the weakness of the classical method, a novel feature extraction method based on features fusion is developed in this paper. First of all, a rational representation for integrated features by virtue of complex vectors is given, i.e., two sets of feature vectors of a same sample are combined together by a complex vector. As a result, the feature space becomes a complex vector space rather than a real one. Then, to solve the problem of feature extraction in the integrated complex vector space, the theory of the complex uncorrelated linear discriminant analysis (ULDA) is developed. Based on this theory, a generalized ULDA method is proposed. This method is suitable for feature extraction in the complex integrated feature space. Finally, the proposed method is tested on Concordia University CENPARMI handwritten digit database and NUST603HW handwritten Chinese character database built in Nanjing University of Science and Technology. The experimental results indicate that after feature extraction using the proposed method, the recognition accuracy is increased significantly as well as the dimension of feature vector is reduced largely. Moreover, the experimental results also demonstrate that the proposed feature extraction method based on feature integration is more powerful and more efficient than the classical one.