Kuzushiji-49
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/rois-codh/kmnist
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为Kuzushiji-49，包含了28x28像素的日本平假名草书字符图像，特别关注字符“あ”（a）和“め”（me），这两个字符因视觉上的相似性而被选中。该数据集对于没有预先了解日文平假名的参与者来说，具有一定的分类难度，且是多模态的，有助于区分不同的字符类别。在规模上，通过亚马逊土耳其机器人招募了488名参与者与该数据集互动，任务是对平假名字符进行二分类。
The dataset is named Kuzushiji-49, which contains 28×28 pixel images of cursive Japanese hiragana characters. Specifically, it focuses on two characters: "あ" (romanized as "a") and "め" (romanized as "me"), which were selected due to their high visual similarity. This dataset poses a non-trivial classification challenge for participants with no prior knowledge of Japanese hiragana, and it is a multimodal dataset that aids in distinguishing between different character categories. A total of 488 participants were recruited via Amazon Mechanical Turk to engage with this dataset, with the core task being binary classification of the hiragana characters.
提供机构：
Kuzushiji-49 authors
