DFCL: Dual-pathway fusion contrastive learning for blind single-image visible watermark removal

Area of Science:

• Computer Vision
• Digital Image Processing
• Machine Learning

Background:

• Digital image watermarking is crucial for copyright protection, leading to research in watermark removal.
• Existing blind visible watermark removal methods face challenges in accuracy and visual quality, often requiring manual mask selection.

Purpose of the Study:

• To introduce a novel dual-pathway fusion contrastive learning approach for blind single-image visible watermark removal.
• To overcome limitations of traditional methods by enhancing feature acquisition and visual quality without manual intervention.

Main Methods:

• A dual-pathway network trained on both image and gradient maps for enhanced feature fusion and spatial positioning.
• Contrastive learning to ensure resemblance to original watermark-free images and separation from watermark content.
• A blind watermark removal algorithm that does not require additional watermark images or mask regions.

Main Results:

• The proposed method significantly improves watermark detection accuracy and spatial positioning.
• Enhanced background visual quality is achieved by effectively removing watermarks.
• The algorithm demonstrates superior performance compared to existing methods on benchmark datasets.

Conclusions:

• The dual-pathway fusion contrastive learning approach offers an effective solution for blind visible watermark removal.
• This method advances the field of digital image copyright protection by improving watermark removal techniques.
• The algorithm provides a robust and efficient tool for recovering watermark-free images.