Seam-carving Localization in Digital Images

Seam-carving is a relatively new image re-targeting technique. While it can be used for legitimate image re-targeting, it also provides a tool for malicious purposes, such as object removal. However, existing methods either classify images in blocks or try to learn faint seam traces from forgeries directly, with low accuracy in the former and inefficiency in the latter. To break these limitations, a new seam-carving localization method is proposed in this research, which can be used to solve the correlation forensic authentication tasks based on seam-carving. JPEG compression brings regular block artifacts to the image,which can be a suitable medium for seam-carving localization. Therefore, we design a multi-block network structure and propose an effective training strategy to localize seams in images. First, we extract the block artifacts hidden in the image self-supervised; second, we input the location map of the seams as guidance to localize the seams from the corrupted properties. As expected, the network can quickly localize the seams with a small amount of training data. By utilizing this prior, we achieve the detection of object removal based on seam-carving. Extensive experiments demonstrate the feasibility and effectiveness of the proposed method.