Fake war crime image detection by reverse image search

Abstract

In the media, images of war crimes are often shared, which in reality come from other contexts or other war sites. In this paper an approach is proposed to detect duplicate or fake war crime images. For this, the bag of visual words model is used in conjunction with localized soft assignment coding and the k-nn classifier. For evaluation, a data set with 600 images of war crimes was crawled. Different distances and parameters were used for evaluation. Unmodified images can be recognized with this approach with 100% accuracy. Rotated and scaled images can also be detected with nearly 100% accuracy. Modifications like cropping or the combination of scaling and cropping ensure significantly smaller accuracy results. The run time was investigated and it was found that about 3000 images per second can be processed on an Intel Core i5 processor.