A self-supervised network for image denoising and watermark removal.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Supervised watermark removal methods require reference non-watermark images, which are often unavailable in real-world scenarios.
• Digital images frequently suffer from noise, complicating watermark removal processes.
• Existing techniques struggle with the dual challenge of denoising and watermark removal without paired data.

Purpose of the Study:

• To develop a self-supervised network (SSNet) for effective and simultaneous image denoising and watermark removal.
• To address the limitations of supervised methods by eliminating the need for reference non-watermark images.
• To create a blind image restoration model applicable to real-world scenarios.

Main Methods:

• A parallel network architecture (SSNet) employing self-supervised learning for denoising and watermark removal.
• Utilizing two sub-networks: an upper network for sequential noise and watermark removal, and a lower network for simultaneous learning.
• Incorporating an attention mechanism between sub-networks to extract complementary salient information.

Main Results:

• The proposed SSNet effectively removes both noise and watermarks without requiring paired training data.
• The method demonstrates superior performance compared to popular image watermark removal techniques on public datasets.
• The self-supervised approach enables blind denoising and watermark removal, enhancing practical applicability.

Conclusions:

• SSNet offers a significant advancement in image restoration by enabling unsupervised denoising and watermark removal.
• The network's ability to handle noise and watermarks simultaneously without reference images makes it highly valuable for practical applications.
• The self-supervised learning strategy and attention mechanism contribute to the method's effectiveness and robustness.