AUTNTdataset
收藏资源简介:
AUTNT数据集包含组件级别的多种脚本文本和非文本图像。该数据集有三个主要用途:(i) 组件级图像分类,(ii) 脚本识别,(iii) 字符识别。数据集分为两类:复杂文档级文本组件和场景级文本组件,以及非文本组件。训练和测试集按照5:1的比例划分,包含多种脚本如拉丁文、孟加拉文和梵文。
The AUTNT dataset encompasses a diverse array of script texts and non-textual images at the component level. This dataset serves three primary purposes: (i) component-level image classification, (ii) script recognition, and (iii) character recognition. It is categorized into two main types: complex document-level text components and scene-level text components, alongside non-text components. The training and testing sets are divided in a 5:1 ratio, featuring a variety of scripts including Latin, Bengali, and Sanskrit.
数据集概述
名称: AUTNTdataset
描述: AUTNTdataset包含多脚本文本和非文本图像,主要用于组件级图像分类、脚本识别和字符识别。数据集中的组件图像来自多种来源和条件,确保了实际应用中的适用性。
数据集组成:
- 总图像数: 10771
- 文本图像数: 7890
- 非文本图像数: 2881
数据集划分:
- 训练集: 6314文本图像,2305非文本图像
- 测试集: 1576文本图像,576非文本图像
文本组件详细信息:
- 文档类型:
- Latin: 训练集1258,测试集314,总计1572
- Bengali: 训练集1002,测试集251,总计1253
- Devanagari: 训练集1004,测试集250,总计1254
- 场景类型:
- Latin: 训练集1759,测试集439,总计2198
- Bengali: 训练集1011,测试集251,总计1262
- Devanagari: 训练集280,测试集71,总计351
非文本组件详细信息:
- 总数: 2881
- 训练集: 2305
- 测试集: 576
相关论文
- T. Khan, A. F. Mollah, “AUTNT - A component level dataset for text non-text classification and benchmarking with novel script invariant feature descriptors and D-CNN”, Multimedia Tools and Applications, vol. 78, no. 22, pp. 32159–32186, 2019.
基准测试结果
文本非文本分类:
- 文档类型:
- Precision: 0.990
- Recall: 0.973
- F-Score: 0.981
- Accuracy: 97.84%
- 场景类型:
- Precision: 0.981
- Recall: 0.931
- F-Score: 0.955
- Accuracy: 95.14%
- 文档和场景类型组合:
- Precision: 0.987
- Recall: 0.961
- F-Score: 0.974
- Accuracy: 96.28%
脚本识别:
- 文档类型:
- Precision: 0.9239
- Recall: 0.9157
- F-Score: 0.9170
- Accuracy: 92.02%
- 场景类型:
- Precision: 0.8139
- Recall: 0.7940
- F-Score: 0.8038
- Accuracy: 89.49%
- 文档和场景类型组合:
- Precision: 0.9149
- Recall: 0.9193
- F-Score: 0.9169
- Accuracy: 92.51%
贡献者
- Tauseef Khan
- Rahamatulla
- Munsi Md Iftabudin
- Mst Fatema Rahman
- Sk Shamim
- Dr. Ayatullah Faruk Mollah




