Research On Image Tampering Detection Algorithm Based On Multi-scale Features And Attention Mechanism

With the rapid development of the Internet and mobile communication devices,users can easily access,edit,and upload images.Although this significantly enriches the online environment in daily life,it also provides a convenient means for malicious individuals to tamper with images and spread them.Image tampering can maliciously guide the spread of negative public opinion in social networks,and can even cause harm to the political stability,economic development,and academic fields of society.Therefore,it is necessary to vigorously develop work to detect image tampering.Existing detection models mostly focus on detecting one or a few specific tampering methods,lacking robust generalization models that can detect all tampering methods.Moreover,existing models often generate false alarms for real images without tampering,greatly limiting the application space of tampering detection models in real life.In recent years,deep learning has made rapid progress,among which convolutional neural networks have performed well in efficient extraction of image features,bringing new breakthroughs to the task of image tampering detection.Unlike tasks such as semantic segmentation and object detection that have made significant progress,tampering detection focuses more on the location of image manipulation areas rather than the semantic content of the image,so this problem should be given more attention from the perspective of algorithm design.In order to further improve the accuracy of detection and localization,this article proposes a dual-branch network structure and a detection and localization strategy from coarse to fine based on multi-scale features and attention mechanism,and adds classification of real images to reduce false positives.The specific research contents are as follows:1.In response to the low detection accuracy of image tampering detection caused by the large differences in tampering features of different tampering operations,the large span of tampering area sizes,and poor network model generalization capabilities,this article proposes a Multi-scale Feature Fusion Network(MFF-Net)for pixel-level tampering area localization.MFF-Net aims to improve the detection accuracy of tampering detection algorithms and effectively detect all tampering methods.It gradually locates the main tampering area through progressive interaction and fusion of different scale resolution information,and introduces image noise feature information that is different from the image RGB information to assist in locating the tampering area of the image by capturing the semantic-independent noise feature in the tampered image.To better integrate the two different features,MFF-Net adopts an attention feature fusion mechanism,which can enhance the information interaction between different branch features and increase the fusion efficiency.The experimental results show that the detection algorithm combined with different scale information can significantly enhance the model’s detection accuracy.The effectiveness of each module was demonstrated through ablation experiments.2.In image tampering detection work,most pixel-level detection methods assume that the input image has already been tampered with before locating the tampered area,and often generate false positives for real images that have not been tampered with.To address this issue and further improve the detection ability of the existing network structure,this article proposes a Reverse Attention Receptive Network(RAR-Net)for tampering detection work.Before tampering detection work,the detector that mixes real and tampered images is used to classify the image to prevent false positives for real images.The feature information of different scale receptive fields is extracted to coarsely locate the tampered area.At the same time,a reverse attention mechanism is used to highlight the tampered area and suppress the interference of irrelevant areas in the image.The experimental results show that the RAR-Net model can effectively detect and locate tampered areas in images and reduce false positives for real images.