One improvement to two-dimensional locality preserving projection method for use with face recognition

While locality preserving projection (LPP) is directly applicable to only vector data, two-dimensional locality preserving projection (2DLPP) is directly applicable to two-dimensional data. As a result, 2DLPP is computationally more efficient than LPP. On the other hand, when determining the transform axes, both conventional 2DLPP and LPP do not exploit the class label information of training samples, the use of which is usually advantageous for producing good classification result. In order to exploit the class label information, we proposed one novel LPP method, i.e. two-dimensional discriminant supervised LPP (2DDSLPP). We also analyzed the characteristics and advantages of 2DDSLPP and presented the difference and relationship between 2DDSLPP and other methods. Compared with two-dimensional discriminant LPP (2DDLPP), 2DDSLPP has a stronger capability to preserve the distance relation of samples from different classes. We used two face databases to test 2DDSLPP and several other two-dimensional dimensionality reduction methods. Experimental results show that 2DDSLPP can obtain a higher classification right rate.