# Towards Photo-Realistic Visible Watermark Removal with Conditional Generative Adversarial Networks

**Authors:** Xiang Li, Chan Lu, Danni Cheng, Wei-Hong Li, Mei Cao, Bo Liu, Jiechao Ma, Wei-Shi Zheng  
**arXiv:** [1905.12845](https://arxiv.org/abs/1905.12845)  
**Date:** 2019-05-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1905.12845v1/

## Key points

- Visible watermarking is important for protecting image copyright
- Robustness of visible watermarks to attacks is essential
- Watermark removal from images has become a hot research topic
- Current methods use encode-decode architectures with pixel-wise loss to remove transparent watermarked pixels
- Existing methods struggle with reconstructing hidden information obscured under complex and various watermarks, resulting in residual watermark traces and unrealistic reconstructed images
- A new watermark processing framework using conditional generative adversarial networks (cGANs) for visible watermark removal in real-world applications is proposed in this paper
- The method uses a patch-based discriminator conditioned on the watermarked images that is adversarially trained to differentiate between recovered images and original watermark-free ones, producing more photo-realistic and convincing results compared to state-of-the-art methods.
- Extensive experimental results on a large-scale visible watermark dataset demonstrate the effectiveness of the proposed method.
- This study provides an innovative solution for improving visible watermark removal in real-world applications using cGANs.

## Summary

The protection of image copyright through visible watermarking has become increasingly important, and the robustness of a visible watermark to attacks is essential. Removing watermarks from images has also gained attention as a hot research topic. Current methods treat watermark removal as an image-to-image translation problem using encode-decode architectures with pixel-wise loss to transfer transparent watermarked pixels into unmarked ones. However, when presented with realistic images, the watermarks are often unknown and diverse, making it difficult for existing methods to reconstruct hidden information obscured under complex and various watermarks. This results in residual watermark traces and reconstructed images that lack reality. To address this difficulty, a new watermark processing framework using conditional generative adversarial networks (cGANs) for visible watermark removal in real-world applications is proposed in this paper. The method enables the watermark removal solution to be more closed to photo-realistic reconstruction by using a patch-based discriminator conditioned on the watermarked images that is adversarially trained to differentiate between recovered images and original watermark-free ones. Extensive experimental results on a large-scale visible watermark dataset demonstrate the effectiveness of the proposed method compared with state-of-the-art methods. The approach produces more photo-realistic and convincing results while addressing the challenges posed by unknown and diverse watermarks in realistic scenarios. Overall, this study provides an innovative solution for improving visible watermark removal in real-world applications using cGANs.