Evaluation of Handwritten Urdu Text by Integration of MNIST Dataset Learning Experience

Abstract

The similar nature of patterns may enhance the learning if the experience they attained during training is utilized to achieve maximum accuracy. This paper presents a novel way to exploit the transfer learning experience of similar patterns on handwritten Urdu text analysis. The MNIST pre-trained network is employed by transferring it's learning experience on Urdu Nastaliq Handwritten Dataset (UNHD) samples. The convolutional neural network is used for feature extraction. The experiments were performed using deep multidimensional long short term (MDLSTM) memory networks. The obtained result shows immaculate performance on number of experiments distinguished on the basis of handwritten complexity. The result of demonstrated experiments show that pre-trained network outperforms on subsequent target networks which enable them to focus on a particular feature learning. The conducted experiments presented astonishingly good accuracy on UNHD dataset.

Keywords

• Transfer learning
• deep analysis
• cursive handwritten data
• MDLSTM