一种基于多尺度内容感知的图像篡改定位方法OA

With the rapid advancement of image editing technologies,image manipulation localization is facing increasingly complex challenges.Existing methods exhibit limitations in capturing multi-scale contextual information of manipulated regions,and they of-ten rely on fixed sampling kernels during the multi-scale feature fusion process,which makes it difficult to focus on local variations of features,thereby restricting the localization accuracy.To address these issues,we propose an image manipulation localization method based on multi-scale content-aware.Firstly,a feature fusion module based on multi-scale content-aware is designed,which can dynamically generate adaptive sampling kernels for each position in the feature map,enabling the model to locate the approxi-mate range of manipulated regions at a coarse-grained level and identify manipulated edge features at a fine-grained level.Secondly,a depthwise separable convolutional decoder is employed to replace the traditional multilayer perceptron for prediction,further en-hancing detection accuracy.Finally,a joint loss function combining binary cross-entropy loss and Dice loss is proposed,effectively improving the model's robustness and generalization capabilities.Cross-dataset experimental results on multiple public datasets dem-onstrate that,in terms of Pixel-level F1 score,the proposed method achieves 69.4％,25.4％,37.8％,and 83.2％on the CASIAv1,De-facto-12k,Coverage,and Columbia datasets,respectively.Compared to the mainstream MVSS-Net++and the latest IML-ViT,it achieves average improvements of 10.2％and 2.1％,respectively,significantly enhancing the accuracy of image tampering localiza-tion.