OCR4JkanjiCards: Exploring Japanese Character Recognition Oswaldo Bassani Neto, Celso Setsuo Kurashima, Marcelo Zanchetta do Nascimento Universidade Federal do ABC {oswaldo.bassani, celso.kurashima, marcelo.zanchetta}@ufabc.edu.br Abstract The OCR4JkanjiCards system is presented in this work, as a proposal for taking photos of Japanese characters using a smartphone and have them automatically recognized within a digital Japanese dictionary. Image processing techniques and neural network strategies are necessary, as well as the exploration of several programming tools. 1. Introduction The increase in processing power and memory available on mobile devices like cell phones, smart- phones or handheld, allow user to have pocket dictionaries providing fast access to information. The goal of this work is to explore a set of tools that allows joining several features in a unique system. Our OCR4JkanjiCards system is based on character recognition, and on the processing of photos taken from street plates, building facades and other texts in the real world. 2. System Overview The OCR4JkanjiCards system basically works with an image of a kanji as the input and gives back the kanji identification as the output, through a user interface as shown in Figure 1. Kanji’s are the complex characters used in Japanese language. Figure 1: OCR4JkanjiCards User Interface. The inner window in the upper left part of Figure 1 is a picture, taken with a cell phone camera, of a poster in the street containing a kanji character. After several image processing steps the kanji was clearly separated in the upper right inner window of Figure 1. In the lower part of this figure, the result of recognition process is shown in the line marked by the green box, in the form of ID codes of kanji’s. These codes refer to a Japanese dictionary where we can obtain the meanings in English. In this case, we used the “Java Kanji Flashcards 500” [1] data base 1 . We created OCR4JkanjiCards system based on the project called OCR4J [2]. This is an open source library that allows the creation of applications with character recognition like OCR (Optical Character Recognition), but is limited to A-Z characters. So our system started from the integration of the OCR system with the Kanji education software, and went to the development of image processing techniques for Kanji characters. We are interested in investigating how kanji’s can be efficiently processed for recognition tasks, given that the character writing is complex and composed of non-connected lines. Correlated work can be found in references [3] and [4]. The first one is a project of kanji dictionary for mobile devices. 3. Methodology Since the purpose of this experiment is the use of images acquired from photographic cameras, we apply some image processing techniques in order to convert input images into binary image format before using the OCR system. So, the initial step is to convert the input colored image to gray scale and to apply a median filter. Then, it is converted to a binary image, with an appropriate threshold. Finally, this image is passed through an 1 "Java Kanji Flashcards 500", available in: http://nuthatch.com/java/kanjicards/, accessed in March 2008.