Image copy detection via learning from copies based on virtual prior attacks

In the near past, most of the researches on image copy detection focused on finding an image feature that is robust for all kinds of image attacks. However, it is difficult to find a feature that is robust enough for most of image attacks. When original image suffers from image attack, especially serious attack, the number of keypoint matches and the performance of image copy detection will be reduced. In this paper, we propose a novel copy detection framework to solve the problem instead of finding a better feature. In our approach, the original image is applied some kinds of virtual attacks, and the attacked images can be considered to be the copies of original image, then we detect keypoint on original image, its copies and test image using scale-invariant feature transform(SIFT) algorithm^[1]. SIFT descriptors are computed for keypoint´s representation. In the past, popular copy detection method used to match the descriptors attached to the keypoints between original image and test image directly. However, our method not only matches directly between original image and test image, but also matches indirectly. This indirect matching process can be converted to two sub-processes: (1) matching the keypoints between test image and original image´s copies, (2) matching the keypoints between original image and its copies. Our method can achieve good performance on keypoint matching and copy detection under the noise and JPEG attack.