Explainable Tampered Text Detection (ETTD)
收藏资源简介:
ETTD数据集是由华南理工大学和蚂蚁集团联合创建的,旨在支持可解释的篡改文本检测任务。该数据集包含21000张图像,其中包括11000张经过篡改的文本图像和10000张真实文本图像,涵盖多语言卡片、文档和场景文本等多种场景。数据集通过多种篡改方法(如复制移动、拼接和生成文本编辑)生成,并使用Poisson Blending技术减少视觉不一致性。数据集的创建过程包括从互联网和现有数据集中收集图像,进行文本篡改,并使用GPT4o生成异常描述。ETTD数据集主要应用于信息安全领域,旨在解决文本图像篡改检测中的黑箱问题,提供可靠的预测和解释。
The ETTD dataset was jointly created by South China University of Technology and Ant Group, aiming to support the task of explainable tampered text detection. This dataset contains 21,000 images in total, including 11,000 tampered text images and 10,000 authentic text images, covering various scenarios such as multilingual cards, documents, and scene texts. The dataset is generated via multiple tampering methods, including copy-move, splicing, and generative text editing, and uses Poisson Blending technology to reduce visual inconsistencies. The dataset creation process includes collecting images from the internet and existing datasets, conducting text tampering, and generating anomaly descriptions with GPT-4o. The ETTD dataset is primarily applied in the field of information security, with the goal of addressing the black-box problem in text image tampering detection and providing reliable predictions and explanations.

- 1Explainable Tampered Text Detection via Multimodal Large Models华南理工大学 · 2024年



