CRFormer: Multi-scale contrastive regularization transformer for visible watermark removal

Visible watermarks are widely used for image copyright protection, but their removal remains a challenging restoration task due to the diversity of watermarks in color, scale, transparency, and spatial distribution. Existing methods predominantly rely on CNN-based frameworks, where limited receptive fields constrain spatial modeling, and contrastive learning is applied at intermediate feature levels rather than directly supervising the final reconstructed output. In this paper, we propose CRFormer, a single-stage Transformer network for blind visible watermark removal. CRFormer replaces the CNN backbone with a full Transformer to model watermark regions of arbitrary spatial extent. To compensate for the reduced spatial sensitivity of channel-wise attention, we introduce a deformable convolution feed-forward network that restores spatial perception and integrates watermark mask prediction directly into the backbone. We further apply contrastive learning as an output-level regularization, where multi-scale VGG features of the reconstructed image are pulled toward the watermark-free ground truth and pushed away from the watermarked input, providing direct supervision over perceptual reconstruction quality that intermediate-feature decoupling cannot offer. Extensive experiments on multiple public benchmarks demonstrate that CRFormer consistently outperforms existing state-of-the-art methods by a significant margin.