二维线性鉴别分析（２DLDA）是一种直接基于矩阵的特征提取方法，跳过传统的基于Fisher鉴别准则 的线性鉴别分析方法中必须先将二维矩阵转化成一维矢量的过程，有效地提高了特征提取速度且避免了小样本问题，其识别率优于传统的Fisherface方法。结合模糊集理论，提出了一种新的２DLDA算法——模糊２DLDA （F１DLDA）算法。首先采用FKNN算法得到相应的样本分布信息，并按其对最后得到的特征向量所作的贡献融入 到特征抽取过程中，得到有效的样本特征向量集。实验表明，F２DLDA算法的性能优于传统的２DLDA算法和Fisherface算法。

2DLDA algorithm is based on2D matrices and overleaps the step of transforming the matrices into the corresponding vectors,which is done on conventional LDA algorithm.However,performance of recognition rate may always be degraded by the overlapping(outlier)samples et al in the field of pattern recognition.How to avoid these shortcomings and extract optimal features to improve the performance of recognition is a key step. In this paper,a new2DLDA algorithm,named fuzzy2DLDA,is proposed.Fuzzy k-nearest neighbour(FKNN) is implemented first to achieve the distribution information of original samples represented with fuzzy membership degrees and is incorporated into the process of feature extraction.The proposed algorithm inherits the virtue of conventional2DLDA and suppresses the shortcoming resulted by overlappin g(outlier)samples et al. Experimental results on AT&T face database demonstrate rec ognition rates of the proposed algorithm outperform that of conventional2DLDA and fisherface.

郑宇杰（1977－），男，浙江舟山市人，南京理工大学博士生