Research On Image Tampering Localization Algorithms Based On Attention Mechanism And Multi-scale Information

In the era of explosive growth of multimedia technology,high-performance electronic products and online programs have brought more opportunities and challenges to the vigorous development of society.People can pick up their mobile phones anytime and anywhere to record their daily life and what they see and hear.Digital image is a kind of information media that is easy to store and widely spread,but illegal elements often use malicious means to tamper with its content,and attempt to deceive the public about public events through false images.In the actual scene,the public needs to quickly judge the tampered area of the forged image.Therefore,image tamper localization algorithm came into being.Although most of the current researches have been able to locate the fake objects of tampered images,they are too dependent on the semantic content and the feature information of the spatial domain,which will affect the algorithm’s judgment of tampered regions and real regions.Image tampering location needs to further explore deeper forgery traces.Based on the attention mechanism and multi-scale information,this paper locates the forged area of tampered images with changed content,and deeply excavates the contour information and frequency characteristics of tampered edges.The specific research contents of this paper are as follows:1.Image tampering localization algorithm based on multi-scale boundaryinteraction learning networkMost of the earlier image tampering localization algorithms lacked strategies for sensing global features and refining edges.In this paper,a multi-scale boundaryinteraction learning network(MB-Net)for image tampering localization is proposed to solve these two problems.Specifically,the adjacent-scale mutual module(ASMM)enhances the perception of contextual information and effectively decreases noise interference when features of different resolutions are fused.In addition,to effectively suppress the model from extracting semantic content,the boundary pixel disparity module(BPDM)computes the convolution of edge pixel disparity at five specific angles to obtain the boundary artifact recognition in the tampered region.The fusion attention module(FAM)aims to integrate information from multi-scale features and edge features and to obtain spatial and channel correlations.Ablation experiments demonstrate that all three modules and edge pixel difference convolution can effectively improve model capability.At the same time,the performance of MB-Net on public standard datasets is no less impressive than that of current mainstream methods.2.Image tampering localization algorithm based on multi-scale frequencyperception enhancement networkCurrent methods for image tampering localization are usually based on the identification of semantic content in the spatial domain.However,the true and false regions of forged images in real scenes lack significant semantic differences,so these algorithms can misjudge the complete tampering cue.To address the above problem and to exploit the frequency difference between the real and tampered regions in tampered images,this paper proposes a frequency-perception network with adversarial training(FP-Net)based on adversarial training for image tampering localization.The frequency separation sensing module(FSSM)refines the high-frequency noise features in the forgery region and retrieves low-frequency information in the spatial domain by separating multi-scale frequency domain features.The global frequency attention module(GFAM)uses extensive sampling and convolution to interactively learn dualdomain frequency information and improve specific perception of the physical location of the falsified region.A small amount of interference noise is fed into the tampered image using adversarial training to force the model to undergo additional training to identify the tampered regions.Experiments on five standard public datasets show that FP-Net is very general,robust and scalable.Compared to other baseline models,FPNet is able to show better localization capabilities.The experimental results show that both models exhibit excellent detection performance in locating tampering traces,and the attention mechanism and multi-scale information can effectively enhance the recognition of tampered images and the localization of tampered regions by the models.