Writer Identification for Offline Handwritten Kanji Characters Using Multiple Features

TLDR: It is shown that a combination of two global features, two local features, and majority voting as a recognizer is efficient for writer identification of offline handwritten Kanji characters.

Abstract:  Abstract—This paper presents a study on character features and recognizers used for writer identification of offline handwritten Kanji characters. It is shown that a combination of two global features, two local features, and majority voting as a recognizer is efficient for writer identification. We performed experiments using an offline Kanji character database containing one-hundred Kanji characters, each written by one-hundred writers, and fifty samples of each Kanji character for a given writer. The experimental results show that the identification rate is 7 points higher than the conventional method using a single feature and obtained an identification rate higher than 99% by using three character classes. Writer identification based on scanned images of handwritten characters is a useful biometric modality with applications in forensic and historical document analysis. Research on writer identification that uses online characters is widespread, but offline characters lack form for conveying dynamic information. Nevertheless, research on writer identification using offline characters has proposed many features to acquire useful information. We studied efficient character features and recognizers for writer identification of handwritten offline Kanji characters. Kanji consists of logographic Chinese characters adopted in Japanese writing. The text-dependent writer identification in our research uses a character recognition process before writer identification because, in text-dependent writer identification, a character class is assumed to be already recognized. Therefore, most of the research on text-dependent writer identification of handwritten characters has the following characteristics: employs features developed for character recognition, does not use multiple character features, and uses the local features of a character. In this paper, we propose efficient character features and a recognizer for text-dependent writer identification of a handwritten Kanji character (1).