Japanese Kana and Brazilian Portuguese Manuscript Database.

This work introduces a Japanese Kana and Brazilian Portuguese database (JKBP), a novel database made freely available aimed at supporting the development of manuscript-based recognition tasks. As far as we know, this is the first database created and made available to the research community composed of Portuguese and Japanese manuscripts. The samples were collected from 57 volunteers. Each volunteer provided five manuscript samples written in Brazilian Portuguese, and five more samples written in Japanese Kana. We also describe baseline experiments accomplished on JKBP exploring some of its different possibilities of classification tasks, such as syllabary and writer classification. We conducted the experiments using Speed Up Robust Features (SURF), a texture descriptor which has been successfully used to capture the textural content in this application domain. Support Vector Machine (SVM) was used for classification, given its good performance in other works available in the literature involving similar tasks. The obtained results showed that the zoning process and use of late fusion provide a meaningful increase of the performance in some scenarios. The best accuracy rates achieved were 97.98% and 83.77% on writer identification task using Portuguese and Japanese manuscripts respectively, and 100% in syllabary classification.