The forensic detection of very similar images

In this paper, we describe how current investigative techniques used in the detection of illegal images are starting to fail. We consider a range of solutions which may be applicable, and in particular consider how they may scale as the size of image collections, both those held by law enforcement agencies and potential criminals, increase. We conclude that there is plenty of promise in the automatic detection of near-duplicate copies of images. However, all available methods are, to varying degrees, probabilistic. In the absence of a convincing large image collection, with a known ground truth of near-duplicate containment, it is impossible to make any serious judgement about the best way forward in terms of understanding the delicate balances among scalability, sensitivity, and specificity of detection.