CNN-Based Copy-Move Forgery Detection Using Rotation-Invariant Wavelet Feature

This paper introduces a machine learning based copy-move forgery (CMF) localization method. The basic convolutional neural network cannot be applied to CMF detection because CMF frequently involves rotation transformation. Therefore, we propose a rotation-invariant feature based on the root-mean squared energy using high-frequency wavelet coefficients. Instead of using three color image channels, two-scale energy features and low-frequency subband image are fed into the conventional VGG16 network. A correlation module is used by employing small feature patches generated by the VGG16 network to obtain the possible copied and moved patch pairs. The all-to-all similarity score is computed using the correlation module. To generate the final binary localization map, a simplified mask decoder module is introduced, which is composed of two simple bilinear upsampling and two batch-normalized-inception-based mask deconvolution followed by bilinear upsampling. We perform experiments on four test datasets and compare the proposed method with state-of-the-art tampering localization methods. The results demonstrate that the proposed scheme outperforms the existing approaches.