On-line Sample Generation for In-air Written Chinese Character Recognition Based on Leap Motion Controller

Abstract

As intelligent devices and human-computer interaction ways become diverse, the in-air writing is becoming popular as a very natural interaction way. Compared with online handwritten Chinese character recognition (OHCCR) based on touch screen or writing board, the research of in-air handwritten Chinese character recognition (IAHCCR) is still in the start-up phase. In this paper, we present an on-line sample generation method to enlarge the number of training instances in an automatic synthesis way. In our system, the in-air writing trajectory of fingertip is first captured by a Leap Motion Controller. Then corner points are detected. Finally, the corner points as well as the sampling points between corner points are distorted to generate artificial patterns. Compared with the previous sample generation methods, the proposed method focuses on distorting the inner structure of character patterns. We evaluate the proposed method on our in-air handwritten Chinese character dataset IAHCC-UCAS2014 which covers 3755 classes of Chinese characters. The experimental results demonstrate that proposed approach achieves higher recognition accuracies and lower computational cost.