Persian handwritten character recognition using convolutional neural network
Behnam Alizadehashraf, Samad Roohi · 2017
Persian handwritten character recognition (PHCR) is one of the challenging issues in machine vision. This study aims to investigate the performance of convolutional neural networks (CNN) on PHCR problems. To investigate the performance of CNN methods, a dataset of Persian handwritten characters has been used as ground truth data. The dataset elements converted into images with the size of 64 × 64 picture elements (pixels). To clarify the outperformance of proposed method, it is compared with notable conventional techniques in PHCR problems. The results show that the performance of conventional methods is not impressive on PHCR. In addition to conventional methods, two types of CNN methods have been implemented. Single convolutional neural network (SCNN) has been implemented based on the simple structure of CNN (LeNet-5). Moreover, the bagging paradigm applied on CNN to extend it into ensemble convolutional neural network (ECNN) with a variety of network parameters. The results show that, although the ECNN outperform SCNN (accuracy=97.1%) in accuracy, SCNN can recognize Persian handwritten characters with bearable complexity and fair accuracy (accuracy=96.3%).