nagohachi/jawildtext_cropped

jawildtext_cropped是一个日语场景文本识别裁剪数据集，源自llm-jp/jawildtext数据集。该数据集通过对源图像中的每个多边形文本区域进行透视变换，将其校正为水平对齐的矩形裁剪图像，适用于训练和评估日语场景文本识别模型。它包含108,403个样本，以webdataset格式组织成22个分片，每个样本包括JPEG格式的校正图像和UTF-8编码的转录文本。预处理过程中，排除了空文本或边长小于8像素的多边形，确保了数据质量。数据集遵循Apache 2.0许可证。jawildtext_cropped is a per-polygon scene-text-recognition crop dataset derived from llm-jp/jawildtext. It provides perspective-warped, rectified bounding rectangle crops from quadrilateral text regions in source images, resulting in tight, horizontally-aligned images suitable for training and evaluating Japanese scene-text recognition models. The dataset contains 108,403 samples, organized into 22 shards in webdataset format, each with a JPEG crop and UTF-8 transcription. Preprocessing steps include filtering out empty-text polygons and those with sides less than 8 pixels, and it is licensed under Apache 2.0.