Visible Watermark Removal via Self-calibrated Localization and Background Refinement

Abstract

Superimposing visible watermarks on images provides a powerful weapon to cope with the copyright issue. Watermark removal techniques, which can strengthen the robustness of visible watermarks in an adversarial way, have attracted increasing research interest. Modern watermark removal methods perform watermark localization and background restoration simultaneously, which could be viewed as a multi-task learning problem. However, existing approaches suffer from incomplete detected watermark and degraded texture quality of restored background. Therefore, we design a twostage multi-task network to address the above issues. The coarse stage consists of a watermark branch and a background branch, in which the watermark branch self-calibrates the roughly estimated mask and passes the calibrated mask to background branch to reconstruct the watermarked area. In the refinement stage, we integrate multi-level features to improve the texture quality of watermarked area. Extensive experiments on two datasets demonstrate the effectiveness of our proposed method. CCS CONCEPTS• Computing methodologies → Image processing.

Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal

“…Several works try to apply neural networks to formulate an end-to-end problem, and there are two popular ways applied to solve it. One way is to directly formulate the watermark removal as an image-toimage translation task [4,29]; the other way is adopting a two-stage strategy to formulate the problem: the first step is to locate the by a mask, and the second step is to recover the background in the watermark area and train a network to solve both at the same time [11,20,32,30]. The latter method was found in experiments can be more effective in watermark removal, so we mainly focus on preventing the second type of networks in this paper.…”

“…Evaluation Metrics. Following the previous work [32,11,30], Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Root-Mean-Square distance (RMSE), and weighted Root-Mean-Square distance (RMSE w ) are adopted as our evaluation metrics. The difference between RMSE and RMSE w is that RMSE w only focuses on the watermarked area.…”

“…Park et al [40] propose to formulate it into a feature matching problem. With the rapid development of deep learning, the community has proposed some blind visible watermark-removal DNNs, which can reconstruct watermarked images end-toend without any information about watermarks [4,8,11,20,29,30,32]. These works usually adopt two-stage strategies.…”

Mask-Guided Image Person Removal with Data Synthesis

“…The inspiration mainly comes from some similar tasks, such as watermark removal [6], [33], raindrop removal [34], and fog removal [35]. These tasks do not or are hard to obtain the object mask m, and the source image I is obtained through α blending, i.e.,…”

“…where G denotes the person removal network, and we use the watermark removal network proposed in [33] as an example. By doing this, we are able to retain the color information to the maximum extent in the process of person removal.…”