A Deep Learning-Based Method for Image Tampering Detection

Abstract

This paper proposes a method to detect the fake objects in images by combining You Only Look Once (YOLO) to define the objects and image post-processing techniques to authenticate the forgery manipulations on objects. YOLO is developed in 2016 and improved in different versions. For the object detection in images, authors used the first version of YOLO. In the proposed method, YOLO is firstly used to identify objects on the image and these objects are removed from the background to reduce computational complexity. The boundaries of objects will be then detected, and their sharpness distribution are calculated as the traces for determining the object tampering. The objects with higher sharpness at the boundaries will be compared the objects in the same group. If there is an object’s features similarity, there will be a copy-move object; otherwise, a spliced object. These steps are integrated into a neural network model trained by both spliced and copy-move image datasets. The combination of YOLO and the proposed neural network model has solved the problem of detecting fake objects with average accuracy of 95.3% for copy-move images and 93.8% for spliced images. The proposed method can be efficient not only for copy-move images and spliced images, but also for the mixed images.