Unsupervised feature learning of Noisy Arabic Characters using wavelet packet transform based SRBM

The unsupervised feature learning is an important task in optical character recognition models. In neural network there are two useful unsupervised architectures; Restricted Boltzmann machine (RBM) and Auto-encoder. In this work, we focus to improve performance the of Stacked RBMs algorithm in order to learn features from a set of Arabic characters in noisy situations. Therefore, we need to update these parameters by the gradient descent algorithm, which minimize cost function without over-fitting. We propose to learn features of noisy Arabic character image using wavelet packet transform before stacking RBMs. This method provides an independent time-frequency information in each level, which helps to minimize overlap of noise in the useful parts. Afterwards, we apply the SRBM and WRBM algorithms on a corpus of Arabic character images, thus, we evaluate their convergence speed by Root Mean Square Error (RMSE) for each RBM algorithm.