Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A self-supervised CNN for image watermark removal

Chunwei Tian, Menghua Zheng, Tiancai Jiao, Wangmeng Zuo, Yanning Zhang, Chia-Wen Lin

TL;DR

This work tackles image watermark removal without relying on ground-truth clean references by introducing SWCNN, a self-supervised CNN architecture. SWCNN combines a self-supervised mechanism to construct reference watermarked images, a heterogeneous U-Net to capture diverse structural information, and a perception-based texture loss to preserve realism, optimized with a mixed L1 loss. Empirical results show SWCNN achieving state-of-the-art PSNR/SSIM across varying watermark transparencies and competitive computational cost, supported by a dedicated watermark dataset with twelve novel watermarks. The approach enhances robustness and practicality for real-world watermark removal scenarios, providing a scalable framework for texture- and structure-preserving image restoration under watermarking threats.

Abstract

Popular convolutional neural networks mainly use paired images in a supervised way for image watermark removal. However, watermarked images do not have reference images in the real world, which results in poor robustness of image watermark removal techniques. In this paper, we propose a self-supervised convolutional neural network (CNN) in image watermark removal (SWCNN). SWCNN uses a self-supervised way to construct reference watermarked images rather than given paired training samples, according to watermark distribution. A heterogeneous U-Net architecture is used to extract more complementary structural information via simple components for image watermark removal. Taking into account texture information, a mixed loss is exploited to improve visual effects of image watermark removal. Besides, a watermark dataset is conducted. Experimental results show that the proposed SWCNN is superior to popular CNNs in image watermark removal.

A self-supervised CNN for image watermark removal

TL;DR

This work tackles image watermark removal without relying on ground-truth clean references by introducing SWCNN, a self-supervised CNN architecture. SWCNN combines a self-supervised mechanism to construct reference watermarked images, a heterogeneous U-Net to capture diverse structural information, and a perception-based texture loss to preserve realism, optimized with a mixed L1 loss. Empirical results show SWCNN achieving state-of-the-art PSNR/SSIM across varying watermark transparencies and competitive computational cost, supported by a dedicated watermark dataset with twelve novel watermarks. The approach enhances robustness and practicality for real-world watermark removal scenarios, providing a scalable framework for texture- and structure-preserving image restoration under watermarking threats.

Abstract

Popular convolutional neural networks mainly use paired images in a supervised way for image watermark removal. However, watermarked images do not have reference images in the real world, which results in poor robustness of image watermark removal techniques. In this paper, we propose a self-supervised convolutional neural network (CNN) in image watermark removal (SWCNN). SWCNN uses a self-supervised way to construct reference watermarked images rather than given paired training samples, according to watermark distribution. A heterogeneous U-Net architecture is used to extract more complementary structural information via simple components for image watermark removal. Taking into account texture information, a mixed loss is exploited to improve visual effects of image watermark removal. Besides, a watermark dataset is conducted. Experimental results show that the proposed SWCNN is superior to popular CNNs in image watermark removal.
Paper Structure (13 sections, 17 equations, 10 figures, 6 tables)

This paper contains 13 sections, 17 equations, 10 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Proposed method
  3. Self-supervised Mechanism
  4. Network Architecture of Self-supervised Convolutional Neural Network
  5. Heterogenous Network
  6. Perception Network
  7. Loss Function
  8. Experiments
  9. Conducted datasets
  10. Experimental setting
  11. Experimental analysis
  12. Experimental results
  13. Conclusion

Figures (10)

  • Figure 1: Architecture of SWCNN.
  • Figure 2: Twelve collected watermarks.
  • Figure 3: PSNR of different methods for one epoch and multiple epochs. (a) PSNR of two methods for one epoch , (b) PSNR of two methods for different epochs
  • Figure 4: PSNR of different methods for one epoch and multiple epochs. (a) PSNR of two methods for one epoch, (b) PSNR of two methods for different epochs.
  • Figure 5: Visual images of different methods: (a) Watermarked image (29.21dB), (b) HN (32.18dB) and (c) SWCNN (35.75dB).
  • ...and 5 more figures