Finding duplicate images in biology papers

Duplicated images in biology papers are a possible indicator for plagiarism or data fabrication. A manual detection of such duplicates can be time consuming or even infeasible for huge image collections. In this paper, a semi-automatic duplicate detection approach is proposed. The approach can be used for the detection of duplicates that cover only a fraction of the full image, are transformed (e.g. rotation), occur between images or within single images (i.e. single-image-duplicates). In the proposed approach, single-image-duplicates are detected between sub-images (i.e. sub-figures) based on a connected component approach and duplicates between images are detected via the min-hashing technique. The approach was evaluated on 1.7 million images extracted from biology papers. By application of various filtering methods to remove false positive detections, only a small amount of manual effort was necessary to find 3041 potentially serious duplicates in so far non-retracted papers.