Image Forgery Detection & Localization Using Regularized U-Net

With the rise in digital media and popular image sharing platforms there has been an increase in the manipulation of images through image editing software. Image editing has never been easier because of the readily available and easy-to-use software. This has led to a wave of tampered images flooding the Internet. Traditionally, the human eye could distinguish between an original image and a tampered one, but with editing software developed recently, it has become significantly harder. Broadly Image Forgery can be either Copy-Move, where a region of an image is copied and pasted on another location in the same Image or Image Splicing, here, a section within a specific region of the image is copied and pasted on another region in a different Image. Most of the current methods and algorithms for Image Forgery detection use manually chosen features to identify and localize manipulated portions of the image with some moving towards Deep Learning models. We followed a deep learning approach of using a modified version of the Image Segmentation Model U-Net. The U-Net model was modified by adding regularization. The results were promising with an F1 score of 0.96 on the validation and test sets with the model able to detect and localize forged sections.