基于双树复数小波四元数卷积网络的Copy-move盲取证算法

随着图像编辑软件的普及与完善,使得人们通过Copy-move操作便可伪造图像,而现有的Copy-move盲取证算法很难提取到彩色图像的一致性特征,且结果依赖于手动调节参数,难以定位到准确的篡改区域.为此,利用四元数卷积网络提取彩色图像空间一致性信息和双树复数小波提取图像局部信息的优势,提出了一种基于双树复数小波四元数卷积网络的Copy-move盲取证算法.首先,将图像表示为四元数并输入到四元数卷积网络中,提取彩色图像的颜色一致性特征,并将双树复数小波变换的高频子带与卷积网络的特征图联合学习图像的局部特征.其次,计算特征向量之间的相似性分数.然后,利用卷积网络提取较高分数的特征,定位相似区域,在一定程度上解决了匹配时手动调节参数的问题;并构建了一个仅定位粘贴区域的辅助分支来区分相似区域.最后,融合了相似和粘贴区域得到能够区分复制和粘贴位置的结果.在CoMoFoD和CASIA CMFD数据集上的主客观实验表明,该算法提升了Copy-move盲取证的定位性能.

Copy-move image blind forgery algorithm based on quaternion convolutional neural networks with dual-tree complex wavelet

With the popularity of image editing software, people can forge images by Copy-move. However, existing Copy-move blind forensics algorithms are difficult to extract consistency features of color images with results depending on manually adjusting parameters, which leads to low locate accurate. Therefore, using the advantage of quaternion convolutional network on extracting spatial consistency of color images and double-tree complex wavelet on extracting image local features, Copy-move blind forensics algorithm based on quaternion convolutional network with double-tree complex wavelet is proposed. Firstly, images represented as quaternions were input to quaternion convolutional network to extract consistency features of color images, and the high-frequency subbands of double-tree complex wavelet transform were connected with feature maps of the convolutional network to learn local features. Next, similarity scores in feature vectors are calculated. Then, convolutional network is extracted features with higher scores to locate similar areas, which solves the problem of adjusting parameters during matching. Besides, an auxiliary branch is constructed to locate the tampered areas. Finally, similar and tampered areas are fused to distinguish copy and paste. The subjective and objective experiments on CoMoFoD and CASIA CMFD datasets show that the proposed algorithm improves the performance of Copy-move blind forgery.