Manga Text Segmentation

Manga Text Segmentation数据集是由布宜诺斯艾利斯大学和阿根廷国家路简大学合作创建的，专注于日本漫画中不受约束文本的像素级分割。该数据集包含900张来自Manga109的漫画图像，每张图像都进行了像素级的文本标注，分为非文本、简单文本（位于对话气泡内）和困难文本（位于对话气泡外）三个类别。创建过程涉及手动标注，旨在为深度学习模型提供高质量的训练数据，以解决漫画文本检测和分割的挑战。该数据集的应用领域包括自动漫画翻译和图像修复，旨在通过自动文本检测和分割技术，降低日本漫画文化的语言障碍。

Manga Text Segmentation Dataset was jointly developed by the University of Buenos Aires and the National University of La Matanza, Argentina, focusing on pixel-level segmentation of unconstrained text in Japanese manga. The dataset includes 900 comic images sourced from Manga109, with each image annotated at the pixel level for text, categorized into three classes: non-text, simple text (located within speech bubbles), and difficult text (located outside speech bubbles). The dataset creation process involved manual annotation, aiming to provide high-quality training data for deep learning models to address the challenges of comic text detection and segmentation. Its application fields cover automatic manga translation and image restoration, with the goal of reducing language barriers to Japanese manga culture via automatic text detection and segmentation technologies.