提出一种基于深度神经网络的图像修复取证算法,该算法可通过编码器网络自动提取图像修复遗留的痕迹特征,通过解码器网络预测像素类别,从而判断出图像是否经过修复篡改以及修复篡改的区域。同时,采用特征金字塔网络对解码器网络中的特征图进行信息补充。采用MIT Place数据集作为训练集,UCID数据集作为测试集,对训练集和测试集分别使用了不同的修复篡改算法。实验结果表明,与其他图像修复取证算法相比,所提算法的修复区域定位更精准,处理速度更快,且对不同的修复篡改方法具有较好的稳健性和较强的泛化能力。

A novel image inpainting forensics algorithm based on the deep neural network is proposed, in which the vestigial features can be automatically extracted by the encoder network, the category of each pixel is predicted by the decoder network, and thus whether or not the image is with inpainting and falsification as well as the inpainted and falsified regions can be distinguished. Simultaneously, the feature pyramid network (FPN) is used to supplement the feature map in the decoder network. The MIT Place dataset is used as the training set and the UCID dataset as the testing set. In addition, the different inpainting and falsification algorithms are adopted for the training set and the testing set, respectively. The experimental results show that, compared with the other inpainting forensics algorithms of images, the proposed algorithm has a more accurate inpainting area and a faster processing speed. Moreover, it has relatively good robustness and strong generalization ability against different inpainting forensics algorithms.