U-Net architecture neural network for localization of digital images integrity violation

Abstract

Subject of Research. The paper presents the study of the U-Net architecture neural network applicability to localization problem of image modifications. The implemented method provides the detecting of the modified image and getting a mask of the changed area. Method. The proposed method was based on deep machine learning – a neural network. U-Net neural network architecture was studied. The training dataset was created as a basis for model training with original images and images modified using a graphical editor. The implemented method represents image as a set of pixels. Main Results. The trained model has shown a high level of brightness recognition for image modifications, up to 80 %, and up to 64 % for copy-shift. Practical Relevance. The result can be practically applicable in the forensics for recognition of modified image blocks and for copyright protection.