A Network Model to Detect the Digital Manipulation in Document Images

Abstract

Digital images usage has rapid growth with increase of internet usage, digital document images are commonly used for sharing information or as submission as proof of information in online systems. This project focuses on developing a deep neural network model to detect the digital manipulation in document images that can be deployed on mobile and edge devices. The proposed model is based on semantic segmentation and modified to use two encoder blocks with one block taking noise generated image as input. Noise is generated utilizing the steganalysis rich model filter layer which creates local noise distribution to extract features from noise inconsistency from manipulated regions of the image. In this project the model is trained on a dataset created using splicing and copy-move digital image manipulation techniques. The authentic images are based on the PubLayNet document image dataset. The model is converted into edge device compatible format so it can run prediction on 32-bit processor systems and doesn’t require a high performance compute device connection.