A new iterative synthetic data generation method for CNN based stroke gesture recognition

Training a stroke gesture classifier by using the state-of-the-art Convolutional Neural Network method requires a large sample size to achieve good performance. This becomes a serious problem when users want to add new gestures to the system because adding so many samples is time-consuming and expensive. In this paper, we propose an iterative synthetic data generation method to solve this problem. The method takes in one user-input template gesture which is modeled by Bezier curve and can generate thousands of samples for training. We propose two different modeling approaches so the method can be applied to both mono and multi-stroke gestures. By applying perturbation to the control points, we can obtain enough samples for training. The generation process is carried out in an iterative way, so the variability in different categories of stroke gestures can be balanced. The variability is measured by the dynamic time wrapping method. The proposed method is tested on our own dataset and two published datasets. Our method outperforms methods with fixed generation process and reaches high recognition accuracy.