Improved Fully Convolutional Network for Digital Image Region Forgery Detection

Abstract

With the rapid development of image editing techniques, the image splicing behavior, typically for those that involve copying a portion from one original image into another targeted image, has become one of the most prevalent challenges in our society. The existing algorithms relying on hand-crafted features can be used to detect image splicing but unfortunately lack precise location information of the tampered region. On the basis of changing the classifications of fully convolutional network (FCN), here we proposed an improved FCN that enables locating the spliced region. Specifically, we first insert the original images into the training dataset that contains tampered images forming positive and negative samples and then set the ground truth masks of the original images to be black images. The purpose of forming positive and negative samples is to guide the improved FCN to distinguish the differences between the original images and spliced images. After these steps, we conducted an experiment to verify our proposal, and the results reveal that the improved FCN really can locate the spliced region. In addition, the improved FCN achieves improved performance compared to the already-existing algorithms, thereby providing a feasible approach for digital image region forgery detection.

AHP validated literature review of forgery type dependent passive image forgery detection with explainable AI

“…Pixel wise information is used for training the model, and Sobel filters use the edge's information to identify the manipulated boundaries. Fully convolutional network model (FCNN) [16] is used for image splicing detection. It distinguishes the altering of an image as well as recognizes the forgery of spliced regions.…”