Segmentation and Recognition of Aerial Handwritten Hiragana

This study try to recognize handwritten hiragana in the air. Five or three characters, such as “a-i-u-e-o” or “ya-yu-yo”, are written overlapping in the same area in the air. The trajectory of the fingertip at that time is acquired as 3D time series data using the Leap Motion Controller. The data value is normalized from the minimum of 0 to the maximum of 1 to reduce the effects of character size and writing position. For character recognition, multiple characters are divided into single characters. To do this, division by “movement strokes” is adopted. Characters are divided by finding “movement strokes,” which are the parts that connect characters to the next character, from the character data and deleting them. By learning and identifying whether a character is a “moving stroke” using LSTM (Long Short Time Memory), character division can be performed even for unknown data. In addition, for each divided character, LSTM is similarly used to classify the 46 hiragana. The input multiple characters and the finally recognized multiple characters are evaluated using the Levenshtein distance. The accuracy rate was 52.333% for a Levenshtein distance cost of 0, 3.500% for 1, 36.167% for 2, and 8.000% for 3 or more. The accuracy rate for a Levenshtein distance cost of 0 when manually segmenting characters was 94.278%, so the accuracy of character segmentation needs to be improved.