A new in-air handwritten persian characters recognition method based on inertial sensor position estimation and convolutional neural network

With advances in microelectromechanical systems (MEMS), researchers have now become interested in the systems operating based on inertial signals. In fact, inertial signals have proven useful in different areas due to advances in their manufacturing technology, availability, and inexpensiveness as well as the development of powerful processing methods such as deep learning techniques. Handwritten character recognition (HCR) is among such areas. This paper aimed to design, implement, and evaluate a novel system for the recognition of handwritten Farsi characters extracted from an inertial pen. For this purpose, a wireless inertial pen was designed. Its motion trajectory was then determined by combining the signals of its angular velocity and acceleration and using the concepts of navigation systems such as quaternion in order to estimate the position signals of characters. A convolutional neural network (CNN) was also employed to facilitate the extraction of high-level features and the classification of characters. The position signal was also extracted as an image used for model learning to enhance the classifier efficiency. The experimental results indicated the CNN-6 architecture outperformed the other CNN-n architectures in terms of character classification accuracy. According to the evaluation of the proposed method through test data, character recognition accuracies of Farsi letters and numbers were reported 91.06% and 94.52%, respectively. In comparison with the previous systems, the proposed method managed to improve the recognition of handwritten Farsi characters.