Angle 2DPCA: A New Formulation for 2DPCA.

2-D principal component analysis (2DPCA), which employs squared -norm as the distance metric, has been widely used in dimensionality reduction for data representation and classification. It, however, is commonly known that squared -norm is very sensitivity to outliers. To handle this problem, we present a novel formulation for 2DPCA, namely Angle-2DPCA. It employs -norm as the distance metric and takes into consideration the relationship between reconstruction error and variance in the objective function. We present a fast iterative algorithm to solve the solution of Angle-2DPCA. Experimental results on the Extended Yale B, AR, and PIE face image databases illustrate the effectiveness of our proposed approach.