Fast classification of handwritten digits using 2D-DCT based sparse PCA

We propose to address the handwritten digits recognition (HWDR) problem by using a two-dimensional (2-D) discrete cosine transform (DCT) based sparse principal component analysis (PCA) algorithm for fast classification. The gain of processing speed is achieved by utilizing the ability of 2-D DCT for energy compaction and signal decorrelation. The proposed algorithm was applied to the mixed national institute for standards and technology (MNIST) database of handwritten digits to demonstrate that when incorporated into the conventional PCA, the 2-D DCT helped reduce the dimension of the input data by 75%. As a result of the dimensionality reduction, the proposed algorithm is 35.7% faster for HWDR than the conventional PCA without sacrificing recognition accuracy.