Mask Dataset for: Unconstrained Text Detection in Manga: a New Dataset and Baseline

This is the dataset used in out paper "Unconstrained Text Detection in Manga: a New Dataset and Baseline". It contains 450 images with the text segmentation of images from Manga109 dataset (need to request access to this dataset in order to view original manga image). Pre-processed version of the images is how they were saved straight out of GIMP. These were later processed before using for training. Post-processed version of the images is after automatically removing small connected components and filling small holes. They are also slightly bigger in width/height in order to be multiples of 8. The text is split in 2 colors: black and pink. Text in black represents text we consider easy to recognize, which is mostly when inside a speech bubble. Text in pink represents text we consider harder to detect, such as text in covers, sound effects or text outside speech bubbles. Further details can be found in our paper.

本数据集用于我们的论文《无约束漫画文本检测：全新数据集与基准方法》（Unconstrained Text Detection in Manga: a New Dataset and Baseline）。该数据集包含450张基于Manga109数据集的图像文本分割标注样本，如需获取原始漫画图像，需先行申请Manga109数据集的使用权限。图像的预处理版本为直接从GIMP导出保存的原始结果，后续在用于模型训练前还会进行进一步加工处理。图像的后处理版本则经过了自动移除小型连通分量、填补小型孔洞的操作，同时为使图像宽高尺寸均为8的整数倍，还对其进行了小幅调整。数据集中的文本分为两种颜色类别：黑色与粉色。黑色文本为我们认定易于识别的文本，这类文本大多位于对话气泡内部；粉色文本则为较难检测的文本，例如封面文字、音效文字或对话气泡外部的文本。更多细节可参阅我们的论文。