Research On Image Forgery Localization Algorithm Based On Feature-reinforced Convolutional Auto-encoders Structure

With the rapid development of digital media technology,various powerful image processing software enables users to obtain,shoot and edit images anytime and anywhere.Most people edit images with very different intentions,but malicious image forgery has caused serious negative effects in some important fields such as the military,media,academic publications,and so on.Earlier methods focused on image-level tamper detection to identify whether an image was forged,and only a few efforts focused on pixel-level image forgery localization.Although domestic and foreign researchers put forward various solutions to locate tampered regions in images,most of these approaches focus on just one or two types of forgery and regard them as semantic segmentation tasks.In recent years,deep learning algorithms show strong information extraction ability and bring new development to image forensics.But,different from semantic segmentation to obtain image content information,tamper detection pays more attention to the manipulation region than to semantic content.In order to further improve localization accuracy,this paper adopts fusion or refinement forms to enhance feature information based on convolutional Auto-encoders structure to prevent irrelevant semantic perception and reduce fault localization.In view of the common copymove,splicing,and removal forgery in digital image content changes,the specific research contents are as follows:1.Hybrid features and semantic reinforcement network for image forgery localizationImage forgery detection focuses more on tampering regions than image content of semantic segmentation,it is revealed that wealthier features need to be learned.Moreover,insufficient semantic information causes low efficiency of forgery localization.To address these issues,this paper proposed a Hybrid Features and Semantic Reinforcement Network(HFSRNet)for image forgery localization,which is an encoding and decoding based network.Specifically,LSTM(Long-Short Term Memory)with resampling features has been applied to capture traces from the image patches for finding manipulating artifacts.Consolidated features extracted from rotating residual units are further leveraged to amplify the discrepancy between un-tampered and tampered regions.We then hybridize features from them through a concatenation to further incorporate the spatial co-occurrence of these two modalities.In addition,for achieving the semantic consistency between two same-level features associated across layers,semantic reinforcement is implemented in the decoding stage.HFSRNet is an end-to-end architecture that handles multiple types of image forgery including copy-move,splicing,and removal.Experiments on three standard image manipulation datasets(NIST16,COVERAGE,and CASIA)demonstrate that HFSRNet obtains state-of-the-art performance compared to existing models and baselines.2.TSFR-Net: A Two-Stage Feature Refinement Network for Image Forgery LocalizationMalicious content-changing images has been considered a potential threat,deep learning-based models illustrated remarkable performance for image forgery detection in recent years.However,most of them suffer from irrelevance semantics awareness while handling forged images of various manipulation categories,leading to false alarms on forged regions.Therefore,image content needs to be suppressed.In this paper,we address aspects by a Two-Stage Feature Refinement Network(TSFR-Net)and follow coarse-to-fine for locating the forged regions given a tampered image.Stage-one,visual attention residual backbone consolidates the learning approach of CNN,which captures an inconsistency between the manipulated and authentic regions straight from the image for subsequent forgery coarse localization.Stage-two,shallow layers features representation strengthened by attention mechanism and then fuse with deep layers features to refine the global forgery features.Next,the coarse localization results guide the model to further fasten on the learning of forged regions rather than semantic content.Finally,the tampering regions are segmented using finer discriminatory features.Experimental results on four standard image manipulation datasets(NIST16,COVERAGE,CASIA v1.0,In-The-Wild)demonstrate that the proposed model performance has improved significantly compared with other state-of-the-art methods.In addition,the robustness comparison of TSFR-Net with another baseline model achieves excellent results on pixel AUC.