Abstract: In order to better deal with the problem of small sample size, and to overcome the defect of two-dimensional locality preserving projection (2DLPP) algorithm which can only keep the local nature of the data, an improved bi-directional two dimensional locality preserving projection algorithm for face recognition is proposed, by combining the characteristics of Two-Dimensional Principal Component Analysis (2DPCA) and Two-Dimensional Linear Discriminate Analysis (2DLDA). First, it introduced the sample class information to improve the weight matrix, and enhances the robustness of the 2DLPP algorithm to samples' changes. Second, two fusion algorithms of 2DLPP+2DPCA and 2DLPP+2DLDA were improved to the feature extraction of row and column direction of the input sample image data. After the feature selection, the optimal projection in row and column direction was obtained. Finally, by performing row and column direction projection on the sample data, the nearest neighbor classification was used to classify the sample data and obtain the recognition results on the given datasets. Experimental results on the face datasets ORL, YALE and AR show that the proposed algorithm is generally superior to the algorithms such as 2DPCA, 2DLDA, 2DLPP, (2D)2PCA, (2D)2LDA, (2D)2PCALDA, and (2D)2LPP-PCA in face recognition performance.