TNCR Dataset (Table Net Detection and Classification Dataset)
收藏资源简介:
我们展示了 TNCR,这是一个从免费开源网站收集的具有不同图像质量的新表格数据集。 TNCR 数据集可用于扫描文档图像中的表格检测并将其分类为 5 个不同的类别。 TNCR 包含 9428 个高质量的标记图像。在本文中,我们实施了最先进的基于深度学习的表格检测方法,以创建几个强大的基线。与其他方法相比,带有 ResNeXt-101-64x4d 骨干网络的 Cascade Mask R-CNN 在 TNCR 数据集上以 79.7% 的精度、89.8% 的召回率和 84.4% 的 f1 得分实现了最高性能。我们将 TNCR 开源,希望能鼓励更多的深度学习方法用于表格检测、分类和结构识别。
We present TNCR, a novel tabular dataset with varying image qualities collected from free and open-source websites. The TNCR dataset can be used for table detection in scanned document images and classifying such tables into 5 distinct categories. TNCR contains 9,428 high-quality annotated images. In this work, we implement state-of-the-art deep learning-based table detection methods to establish several strong baselines. Compared with other methods, Cascade Mask R-CNN with a ResNeXt-101-64x4d backbone achieves the highest performance on the TNCR dataset, with a precision of 79.7%, recall of 89.8%, and F1-score of 84.4%. We open-sourced TNCR with the aim of encouraging further deep learning approaches for table detection, classification, and structural recognition.




