A hierarchical feature learning for isolated Farsi handwritten digit recognition using sparse autoencoder

In recent years, the recognition of Farsi handwritten digits is drawing increasing attention. Feature extraction is a very important stage in handwritten digit recognition systems. Recently deep feature learning got promising results in English handwritten digit recognition, though there are very few papers in this area for Farsi handwritten digits. The contribution of this paper is to propose a new framework utilizing a two layer sparse autoencoder for feature learning directly from data and using the learned weights for feature extraction. In the classification stage of our proposed framework Softmax regression is applied. This recognition method is applied to Farsi handwritten digits in the HODA dataset. The experimental results support our claim that use of deep feature learning as feature extraction stage improves the performance compared with conventional methods.