Efficient two dimensional principal component analysis for online learning
Weiya Shi · 2009
Recently, two dimensional principal component analysis (2dPCA) has attracted much attention, which need not to transform the matrix into vector like traditional PCA. But it has some disadvantage when faced with large-scale data set or online learning, the covariance matrix must be recomputed. In addition, the storage problem also makes its computation impossible. In this paper, an efficient online learning algorithm is proposed to treat with the problem. The essence of two dimensional PCA was firstly analyzed, which states that the column vector or row vector for each image can be treated as the special input vector. Thus, it can be used as input sample for the general incremental PCA algorithm. The proposed method uses less storage and has quick convergence. The effectiveness is demonstrated by the experiment results on the ORL and YaleB databases.