Hierarchical Sparse Autoencoder Using Linear Regression-Based Features In Clustering For Handwritten Digit Recognition

Recently, handwritten digit recognition using higher level features has got more promising results than conventional ones using intensity values, where the higher level features are considered as features of simple strokes in images. Although the state-of-the-art performance is very impressive, there is still room to improve better in both accuracy and computation complexity. In this paper, we propose a new feature based on linear regression to extract geometrical characteristics of handwritten digits. The linear regression-based features are utilized to cluster set of digit image in preprocessing. After that, each set of clustered digit images is inputted a hierarchical sparse autoencoder to extract higher level features automatically. Our method result achieves error rates lower than that of conventional method in the most of cases. The experiment shows that the efficiency of data clustering can get promising results.