Object insertion and removal in images with mirror reflection

Abstract

In this paper, we study reflection integrity assessment for images with strong mirror reflection. Image reflections are physical-level forensic cues involving complicated interactions between surface materials, geometry and lighting, and therefore extremely difficult to fake. Malicious photo manipulations, such as object insertion and removal, can be detected by predicting reflection locations and geometry using scene content and comparing reflections with directly observed objects. We propose a reflection-invariant Bag-of-Features approach to detect and match interest points in the scene and reflection regions, without any prior knowledge. The proposal is open to any robust features and seeks for the right feature yielding the maximal number of matched points. In addition, robust change detection based on disjoint information is proposed to detect object insertion and removal, which is less sensitive to incidental appearance changes. The proposed method is validated on 868 images from the world dataset to demonstrate its efficacy.