Image Inpainting Detection Based On High-pass Filter Attention Network

With the development of the Internet era,digital images are widely used in our daily life,and the popularization and use of image editing tools and deep tampering technology can easily tamper with the target image area with realistic content.The processed image distorts the real information of the image.If it is maliciously used by criminals,such as removing the content in court evidence,removing the watermark visible in copyright,etc.,it will cause serious harm to society.Tampering forensics research based on image inpainting can detect and locate the inpainted area,allowing people to better understand the real content expressed by the image.Therefore,the research of image inpainting and forensics has important social significance and value.Recent research shows that a high-pass filtered fully convolutional network is applied to image inpainting detection and achieves good results.However,those methods did not consider the spatial location and channel information of the feature map.To solve these shortcomings,this paper proposes a squeeze and excitation(SE)block-based high-pass filtered attention fully convolutional network to localize image regions subjected to deep-inpainting operations.The proposed method first uses the high-pass filter to design the initial pre-filter module and extract image residuals to enhance the traces left by the inpainting.Then 4 residual network(ResNet)blocks are used to build a feature extraction module,which learns recognizable features from the high-pass residuals,and uses parallel spatial and channel squeeze and excitation(scSE)block to enhance feature extraction.Finally,an upsampling module is used to predict pixellevel labels of the image,and channel squeeze and excitation(cSE)block is used to enhance the localization of tampered regions,resulting in the final forensic result.The experimental results show that the method can effectively detect and locate the repaired area,which fully confirms the scientific feasibility of the method.Although the attention mechanism can help deep neural networks suppress less significant pixels or channels,existing attention lacks the consideration of the influencing factors of weights,which can further suppress unimportant channels or pixels.Therefore we use the influence factor of the weight to improve the attention mechanism,use the scaling factor of the batch norm alization(BN),and use the standard deviation to express the importance of the weight.High-pass filtered attention fully convolutional network based on Normalization-basedAttention Module(NAM)is proposed,using normalization-based attention module to improve the performance of the previously mentioned high-pass filtered attention fully convolutional network.The experimental results show that the method can effectively detect the repaired region.