IMD2020: A Large-Scale Annotated Dataset Tailored for Detecting Manipulated Images

0524641 - ÚTIA 2021 RIV US eng C - Conference Paper (international conference)
Novozámský, Adam - Mahdian, Babak - Saic, Stanislav
IMD2020: A Large-Scale Annotated Dataset Tailored for Detecting Manipulated Images.
2020 IEEE Winter Applications of Computer Vision Workshops (WACVW). Piscataway: IEEE, 2020, s. 71-80. ISBN 978-1-7281-7162-3.
[2020 Winter Conference on Applications of Computer Vision (WACV ’20). Snowmass Village, CO (US), 01.03.2020-05.03.2020]
EU Projects: European Commission 825227
Institutional support: RVO:67985556
Keywords : Image forensics * Image processing * CNN
OECD category: Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
http://library.utia.cas.cz/separaty/2020/ZOI/novozamsky-0524641.pdf

Witnessing impressive results of deep nets in a number of computer vision problems, the image forensic community has begun to utilize them in the challenging domain of detecting manipulated visual content. One of the obstacles to replicate the success of deep nets here is absence of diverse datasets tailored for training and testing of image forensic methods. Such datasets need to be designed to capture wide and complex types of systematic noise and intrinsic artifacts of images in order to avoid overfitting of learning methods to just a narrow set of camera types or types of manipulations. These artifacts are brought into visual content by various components of the image acquisition process as well as the manipulating process. In this paper, we introduce two novel datasets. First, we identified the majority of camera brands and models on the market, which resulted in 2,322 camera models. Then, we collected a dataset of 35,000 real images captured by these camera models. Moreover, we also created the same number of digitally manipulated images by using a large variety of core image manipulation methods as well we advanced ones such as GAN or Inpainting resulting in a dataset of 70,000 images. In addition to this dataset, we also created a dataset of 2,000 “real-life” (uncontrolled) manipulated images. They are made by unknown people and downloaded from Internet. The real versions of these images also have been found and are provided. We also manually created binary masks localizing the exact manipulated areas of these images. Both datasets are publicly available for the research community at http://staff.utia.cas.cz/novozada/db.
Permanent Link: http://hdl.handle.net/11104/0309167