A neural framework for online recognition of handwritten Kanji characters

Abstract

The aim of this study is to propose an efficient and fast framework for recognition of Kanji characters working in a real-time during their writing. Previous research on online recognition of handwritten characters used a large dataset containing samples of characters written by many writers. Our study presents a solution that achieves fine results, using a small dataset containing a single sample for each Kanji character from only one writer. The proposed system analyses and classifies the stroke types appearing in a Kanji and then recognises it. For this purpose, we utilise a Convolutional Neural Network and a hierarchical dictionary containing Kanji definitions. Moreover, we compare the histograms of Kanjis to solve the problem of distinguishing character having the same number of strokes of the same type, but arranged in a different position in relation to each other. The proposed framework was validated experimentally on online handwritten Kanjis by beginners and advanced learners. Achieved accuracy up to 89 % indicates that it may be a valuable solution for learning Kanji by beginners.