Blind Visible Watermark Removal with Morphological Dilation

TL;DR

The paper tackles blind visible watermark removal without requiring target (watermark-free) images. It introduces MorphoMod, a three-stage pipeline (Segment, Inpaint, Restore) that leverages morphological dilation to refine watermarks and perform diffusion-based inpainting without background priors. It also provides new Alpha1-S/L datasets with opaque watermarks and novel blind evaluation metrics, and demonstrates strong watermark removal performance with controlled trade-offs to preserve semantics, supplemented by ablation studies and a case study on steganographic disorientation. The work advances practical watermark removal in real-world scenarios and opens avenues for future research in image restoration and adversarial manipulation of visible content.

Abstract

Visible watermarks pose significant challenges for image restoration techniques, especially when the target background is unknown. Toward this end, we present MorphoMod, a novel method for automated visible watermark removal that operates in a blind setting -- without requiring target images. Unlike existing methods, MorphoMod effectively removes opaque and transparent watermarks while preserving semantic content, making it well-suited for real-world applications. Evaluations on benchmark datasets, including the Colored Large-scale Watermark Dataset (CLWD), LOGO-series, and the newly introduced Alpha1 datasets, demonstrate that MorphoMod achieves up to a 50.8% improvement in watermark removal effectiveness compared to state-of-the-art methods. Ablation studies highlight the impact of prompts used for inpainting, pre-removal filling strategies, and inpainting model performance on watermark removal. Additionally, a case study on steganographic disorientation reveals broader applications for watermark removal in disrupting high-level hidden messages. MorphoMod offers a robust, adaptable solution for watermark removal and opens avenues for further advancements in image restoration and adversarial manipulation.

Figures (14)

• Figure 1: Example watermark localization masks generated with SLBR liang2021visible, a visible watermark removal method, during the removal process. The predicted masks are conservative compared to the ground truth (GT), leading to watermark residuals in the processed image.
• Figure 2: A diagram of MorphoMod---the proposed automated visible watermark removal method with mask refinement and generative inpainting. MorphoMod consists of three main phases: (1) segment, (2) inpaint, and (3) restore. In segment, an input image $x$ and dilation parameter $d$ are used to produce a refined and dilated mask $\hat{m}_d$. The input image, a prompt $p$, and the refined mask are then used to inpaint the detected region resulting in a cleaned image $\hat{x}$. In restore, this cleaned image, the input image, and the refined mask are used to generate the final restored image with the removed watermark $\dot{x}$.
• Figure 3: An example restoration. A generated image is combined with the refined mask to select only the watermarked region. The input image and the inverse of the refined mask are used to select the background region. The watermarked region and background region are then combined to create the restored image.
• Figure 4: Watermark removal (WR) and semantic preservation (SP) metrics for MorphoMod on the CLWD and LOGO-series datasets across various dilation values $d$.
• Figure 5: Visible watermark removal results on the Alpha1-S dataset. Our method, MorphoMod, is the only method to successfully remove the opaque watermarks.