[Paper Review] WDNet: Watermark-Decomposition Network for Visible Watermark Removal

Yang Liu, Zhen Zhu+1 more|arXiv (Cornell University)|Dec 14, 2020

Advanced Steganography and Watermarking Techniques26 references4 citations

TL;DR

WDNet is a two-stage deep learning framework that decomposes visible watermarks from images by first localizing and estimating rough watermarks and transparencies, then refining the results with a lightweight refinement network. It achieves state-of-the-art performance in watermark removal and separation, with improved generalization through dataset augmentation using separated watermarks, validated on the new CLWD dataset containing 200 colored watermarks.

ABSTRACT

Visible watermarks are widely-used in images to protect copyright ownership. Analyzing watermark removal helps to reinforce the anti-attack techniques in an adversarial way. Current removal methods normally leverage image-to-image translation techniques. Nevertheless, the uncertainty of the size, shape, color and transparency of the watermarks set a huge barrier for these methods. To combat this, we combine traditional watermarked image decomposition into a two-stage generator, called Watermark-Decomposition Network (WDNet), where the first stage predicts a rough decomposition from the whole watermarked image and the second stage specifically centers on the watermarked area to refine the removal results. The decomposition formulation enables WDNet to separate watermarks from the images rather than simply removing them. We further show that these separated watermarks can serve as extra nutrients for building a larger training dataset and further improving removal performance. Besides, we construct a large-scale dataset named CLWD, which mainly contains colored watermarks, to fill the vacuum of colored watermark removal dataset. Extensive experiments on the public gray-scale dataset LVW and CLWD consistently show that the proposed WDNet outperforms the state-of-the-art approaches both in accuracy and efficiency. The code and CLWD dataset are publicly available at https://github.com/MRUIL/WDNet.

Motivation & Objective

• To address the limitations of existing watermark removal methods that lack watermark separation and generalization to diverse watermark styles.
• To develop a deep learning model capable of both removing and explicitly decomposing watermarks from images, enabling reuse of separated watermarks for training data augmentation.
• To construct a large-scale, diverse dataset (CLWD) with 200 colored watermarks to address the lack of real-world colored watermark training data.
• To achieve high performance in both removal accuracy and inference efficiency while avoiding post-processing steps.
• To improve model generalization by leveraging separated watermarks to augment training data.

Proposed method

• WDNet employs a two-stage generator: the first stage uses a U-Net backbone to predict rough watermarks and their transparencies, enabling initial decomposition of the watermark-free image.
• The first-stage output is used to reconstruct a preliminary watermark-free image using the inverse image composition formula.
• The second stage employs a small Refinement network composed of residual blocks to perform pixel-wise refinement on the initial output, focusing on local details.
• The two-stage network is trained jointly without requiring post-processing, ensuring end-to-end optimization.
• The model leverages the ability to separate watermarks from unseen images, which are then used to synthesize new training pairs for data augmentation.
• A new large-scale dataset, CLWD, is constructed with 200 colored watermarks to support training and evaluation on more realistic, diverse watermark patterns.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively decompose visible watermarks from images while achieving high-quality removal without post-processing?
• RQ2 Can the separated watermarks from unseen images be effectively reused to improve model generalization through data augmentation?
• RQ3 Does a two-stage architecture that first localizes watermarks and then refines the result outperform end-to-end image-to-image translation in watermark removal?
• RQ4 How does the performance of a watermark removal model trained on a diverse, colored watermark dataset (CLWD) compare to one trained on a grayscale-only dataset (LVW)?
• RQ5 To what extent does watermark separation enhance the robustness and generalization of watermark removal models?

Key findings

• WDNet achieves state-of-the-art performance on both the LVW and CLWD datasets, outperforming previous methods in all major metrics including PSNR, SSIM, and LPIPS.
• The model demonstrates superior generalization, especially in cross-dataset evaluation, where it maintains strong performance when tested on data from a different distribution (e.g., LVW → CLWD and CLWD → LVW).
• The use of separated watermarks from CLWD-60 to augment training data on CLWD-100 leads to measurable performance gains, confirming the effectiveness of watermark separation for data augmentation.
• WDNet is faster than most SOTA methods, with inference speed second only to Li et al. [16], indicating a good balance between accuracy and efficiency.
• The CLWD dataset contains more diverse and generalized watermark patterns than LVW, as evidenced by the higher performance gap when transferring from LVW to CLWD.
• Visual results show that WDNet produces outputs with the best structural consistency to ground truth images, and successfully separates watermarks that can be reused for training.