Scholarlink

<p>In prediction-based reversible data hiding (RDH), achieving high prediction accuracy is essential for obtaining high-fidelity marked images with increased embedding capacity. To this end, this paper proposes two novel techniques: a self-attention-based convolutional neural network predictor (SA-CNNP) and an error adjustment strategy for color images. The SA-CNNP effectively captures both local characteristics and global pixel dependencies, resulting in comprehensive coverage and improved prediction accuracy. The error adjustment strategy further enhances accuracy by refining the prediction errors of two color channels using the error distribution of a reference channel, thereby promoting inter-channel consistency. Experimental results demonstrate that these innovations lead to significant improvements. More specifically, the proposed SA-CNNP achieves a sharper prediction error histogram with approximately 8<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> improvement in MSE over the best known state-of-the-art predictor. Additionally, the error adjustment strategy increases prediction accuracy for ’Kodim09’ color image by 58<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>. Consequently, the proposed RDH approach achieves an average PSNR gain of around 1.2 dB compared to existing methods for color images.</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

Reversible data hiding for color images using a novel self-attention based CNN predictor and error adjustment

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Sonal Gandhi,
  • Rajeev Kumar

摘要

In prediction-based reversible data hiding (RDH), achieving high prediction accuracy is essential for obtaining high-fidelity marked images with increased embedding capacity. To this end, this paper proposes two novel techniques: a self-attention-based convolutional neural network predictor (SA-CNNP) and an error adjustment strategy for color images. The SA-CNNP effectively captures both local characteristics and global pixel dependencies, resulting in comprehensive coverage and improved prediction accuracy. The error adjustment strategy further enhances accuracy by refining the prediction errors of two color channels using the error distribution of a reference channel, thereby promoting inter-channel consistency. Experimental results demonstrate that these innovations lead to significant improvements. More specifically, the proposed SA-CNNP achieves a sharper prediction error histogram with approximately 8 \(\%\) % improvement in MSE over the best known state-of-the-art predictor. Additionally, the error adjustment strategy increases prediction accuracy for ’Kodim09’ color image by 58 \(\%\) % . Consequently, the proposed RDH approach achieves an average PSNR gain of around 1.2 dB compared to existing methods for color images.