Robust location-secured high-definition image watermarking based on key-point detection and deep learning

Abstract

In this paper, we propose a blind high-definition image watermarking scheme named multiscale-fusion dilated ResNet-based high-definition watermarking (MDResNet-HDWM), which is based on a key-point detection and deep learning framework. The proposed watermarking scheme embeds watermark in multiple non-overlapping regions, whose locations are secured with a private key referring to several dominant key points. To achieve scale-invariant watermark embedding, regions are first determined in a normalized image copy and then mapped back to its origin. The watermarks are embedded in the central part of the region in such a way that the watermarks will always be located inside the regions even if the image is geometrically transformed. The watermarks are embedded and extracted with MDResNet, which is trained with a curriculum learning strategy that makes it be robust to signal processing operations and geometric transforms. Experimental results demonstrate that the proposed MDResNet-HDWM achieves good performance and is robust to both common signal operations and geometric attacks.