SurFITR (Surveillance Forgery Image Test Range)
收藏资源简介:
SurFITR是由墨尔本大学和新加坡管理大学联合构建的监控场景图像伪造检测数据集,包含13.7万张经过局部篡改的监控图像,覆盖人类和物体的删除、替换、添加等精细操作。数据集通过多模态大模型驱动的自动化流程生成,整合了六种主流监控视频源和五种图像编辑模型,提供像素级篡改标注。其创新性体现在语义感知的局部篡改模拟、跨场景/跨模型的组合设计,以及针对监控场景低画质、小目标等特点的针对性构建,旨在解决现有伪造检测模型在监控场景中泛化性不足的问题,推动数字取证技术在安防领域的应用。
SurFITR is a surveillance scene image forgery detection dataset jointly constructed by the University of Melbourne and Singapore Management University. It contains 137,000 locally tampered surveillance images, covering fine-grained manipulations including deletion, replacement and addition of humans and objects. The dataset is generated via an automated pipeline driven by multimodal large models, integrating six mainstream surveillance video sources and five image editing models, and provides pixel-level tampering annotations. Its innovations lie in semantic-aware local tampering simulation, cross-scenario and cross-model combined design, and targeted construction tailored to the characteristics of surveillance scenes such as low image quality and small targets. It aims to address the insufficient generalization performance of existing forgery detection models in surveillance scenarios, and promote the application of digital forensics technology in the security field.
SurFITR 数据集概述
数据集名称
SurFITR (Surveillance Forgery Image Test Range)
核心用途
用于监控风格图像伪造检测与定位。
开发背景
为应对开放访问图像生成模型的最新进展可能带来的伪造视觉证据风险而创建。现有的伪造检测模型在针对监控场景时泛化能力不足,因为监控图像中的篡改通常具有局部性、隐蔽性,且场景视角多变、主体小或被遮挡、视觉质量较低。
数据集内容与规模
- 图像数量:包含超过 137,000 张篡改图像。
- 图像特性:具有不同分辨率和编辑类型的图像。
- 生成方法:通过多模态大语言模型驱动的流程生成,支持跨多样监控场景的语义感知、细粒度编辑。使用了多种图像编辑模型进行生成。
实验验证
广泛的实验表明:
- 现有检测器在 SurFITR 数据集上性能显著下降。
- 在 SurFITR 上训练能显著提升模型在域内和跨域的性能。
许可证
数据集将在 CC BY-NC 4.0 许可证下发布。
引用信息
如需引用,请使用以下格式: bibtex @article{wang2026surfitr, title={SurFITR: A Dataset for Surveillance Image Forgery Detection and Localisation}, author={Wang, Qizhou and Pang, Guansong and Leckie, Christopher}, journal={arXiv preprint arXiv:2604.07101}, year={2026} }
相关资源链接
- 论文地址:https://arxiv.org/abs/2604.07101
- HuggingFace 数据集地址:https://huggingface.co/datasets/wqz995/SurFITR
- 附录地址:https://drive.google.com/file/d/1OaYEI_iWdoCq9JH3tY52qxV9P1Tu7Pvo/view?usp=sharing




