读光-OCR-自然场景信息抽取数据集-中英

SIBR是面向自然场景视觉信息抽取的数据集。相对已有FUNSD、XFUND、CORD等数据集，SIBR的图片具有噪音、打印偏移、重叠、模糊、褶皱等特点，对文字检测、识别、信息抽取都具有挑战。

SIBR is a dataset for visual information extraction from natural scenes. Compared with existing datasets such as FUNSD, XFUND, CORD and others, the images in SIBR feature characteristics including noise, print misalignment, overlapping, blurriness, creases and other issues, which pose significant challenges to text detection, text recognition and information extraction tasks.

搜集汇总

背景与挑战

背景概述

SIBR是一个自然场景视觉信息抽取（VIE）数据集，包含1000张中英文图片，分为600张训练和400张测试。其标注遵循FUNSD和XFUND规范，实体类别包括header、question、answer、other，并标注了inter和intra两种链接类型以构建实体和键值对。