Research Of Image Manipulation Detection Based On Deep Learning

With the development of image editing techniques,people can easily use various manipulation techniques to change the semantic content of images easily.These fake images have become more realistic and gradually threatening the security of image content.In contrast,capturing sparse tampering clues in fake images and completing general image manipulation detection is still challenging.In this situation,this paper attempts to extract more general image tampering features through deep learning technology in a multi-task framework and achieve general image manipulation detection in complex natural environments.The main content of this paper includes:1.Due to the existing digital image manipulation detection model does not match the actual image forensics task.This paper redefines image manipulation detection based on the actual image forensics process and subdivides it into two subtasks: tampering type classification and tampering area localization.According to the definition of image manipulation detection,a new two-stage general image manipulation detection paradigm under the multi-task framework is designed.2.Due to the existing image forensics models are limited by hand-crafted features,and these models lack generalization ability and robustness.This paper proposed a general image manipulation detection model named CRCNN,which utilizes the learnability of the constrained convolutional layer to extract general forgery clues and creates a unified feature representation of various content manipulation with CNN.Subsequently,a twostage general detector gradually guides the model to complete the task of tampering type classification and tampering area localization.Experimental results show that CRCNN can achieve better image manipulation detection performance.3.Due to the existing methods have poor performance on deepfake image.This paper proposes a general image manipulation detection model named MMIG-Net based on the multi-modality image feature interactive mechanism.In MMIG-Net,the RGB modality image features and the frequency modality image features are extracted in parallel through the convolutional neural network.Especially,the frequency modality image is generated by the constrained convolutional layer from the RGB modality image.The model uses the attention mechanism to map the two modality features to a unified vector space as the attention weights.The attention weights of two modalities repeatedly interact to guide the feature extraction process.Subsequently,the model introduces edge supervision signals for the frequency feature stream based on the two-stage detector,which optimizes the effect of frequency modality feature extraction.The experimental results prove that MMIG-Net can effectively deal with various image forgery techniques and complex fake images.4.In order to fill the gaps in deepfake natural scenario images dataset,this paper proposes a deefake images dataset named Synthetic Co Co.The dataset covers manual tampered images,classic computer tampered images and deepfake images.A variety of post-processing methods are used to strengthen the fake samples.The dataset contains49.5K tampered image samples,which meeting the training and testing of current deep learning models.Moreover,SyntheticCoCo providing comprehensive annotation,including manipulation type,tampering techniques names,tampering region mask,tampering boundary mask.