Scholarlink

<p>To solve the problems of existing wavelet-deep learning image watermarking methods lack comprehensive robustness or have high computational complexity, this paper proposes RWWA, a robust watermarking framework that integrates wavelet-transform convolution (WTConv) with attention mechanism unit (AMU). The WTConv expands the receptive field and preserves spatial-frequency details while maintaining parameter efficiency. The AMU enhances robust features through channel-spatial attention mechanisms, while the multi-scale feature extraction (MSFE) modules in the decoder integrate with WTConv to optimize watermark extraction accuracy. Experiments show RWWA outperforms HiDDeN with 93.02% lower BER under attacks, while maintaining acceptable visual quality with only 0.71M parameters. The RWWA prioritizes robustness and computational efficiency while maintaining acceptable visual quality, making it a practical solution for real-world image watermarking applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust watermarking via wavelet-transform convolution and attention mechanism

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Ying Wei,
  • Tianqi Zhang,
  • Pan Deng,
  • Debang Liu,
  • Linhao Cui,
  • Baoze Ma

摘要

To solve the problems of existing wavelet-deep learning image watermarking methods lack comprehensive robustness or have high computational complexity, this paper proposes RWWA, a robust watermarking framework that integrates wavelet-transform convolution (WTConv) with attention mechanism unit (AMU). The WTConv expands the receptive field and preserves spatial-frequency details while maintaining parameter efficiency. The AMU enhances robust features through channel-spatial attention mechanisms, while the multi-scale feature extraction (MSFE) modules in the decoder integrate with WTConv to optimize watermark extraction accuracy. Experiments show RWWA outperforms HiDDeN with 93.02% lower BER under attacks, while maintaining acceptable visual quality with only 0.71M parameters. The RWWA prioritizes robustness and computational efficiency while maintaining acceptable visual quality, making it a practical solution for real-world image watermarking applications.