Multi-scale and deeply supervised network for image splicing localization

When maliciously tampered images are disseminated in the media, they can potentially cause adverse effects and even jeopardize national security. Therefore, it is necessary to investigate effective methods to detect tampered images. As a challenging task, the localization of image splicing tampering investigates whether an image contains tampered regions spliced from another image. Given the lack of global information interactions in existing methods, a multi-scale, deeply supervised image splicing tampering localization network is proposed. The proposed network is based on an encoder–decoder architecture, where the decoder uses different levels of feature maps to supervise the locations of splicing, enabling pixel-wise prediction of tampered regions. Moreover, a multi-scale feature extraction module is utilized between the encoder and decoder, which expands the global view of the network, thereby enabling more effective differentiation between tampered and non-tampered regions. F1 scores of 0.891 and 0.864 were achieved using the CASIA and COLUMB datasets, respectively; and the proposed model was able to accurately locate tampered regions.