Unsupervised Image Manipulation Localization With Non-Binary Label Attribution

Existing forensic techniques for image manipulation localization crucially assume that probe pixels belong to one of exactly two classes, genuine or manipulated. This letter argues that this convention fuels mis-labeling particularly in unsupervised settings, where singular but genuine content or the presence of multiple distinct manipulations may easily induce non-optimal partitions of the feature space. We propose to relax constraints via a greedy $n$ -ary clustering approach, which we instantiate exemplarily in the popular pixel descriptor space of residual co-occurrences. Experimental results on widely used public benchmark datasets highlight the benefits of our approach.