Scholarlink

<p>Deep learning has revolutionized data hiding by automating the determination of hidden positions and intensity settings through network training under specific loss functions. The large capacity of image hiding has garnered considerable attention from researchers. To address the common issues of shadow and contour distortions in multi-image hiding, we propose a novel reversible neural network framework for embedding multiple low dynamic range (LDR) images into a single high dynamic range (HDR) image. This framework, termed IMMIH, leverages the properties of invertible neural networks (INNs) to achieve fully coupled and reversible information hiding and extraction through a single end-to-end training process. By focusing on the RGB channels of HDR images, IMMIH minimizes distortion while incorporating attention mapping and edge mapping modules to improve adaptability. Furthermore, the use of discrete wavelet transform and low-frequency wavelet loss concentrates data hiding in high-frequency regions, ensuring robustness. Experimental results demonstrate that IMMIH achieves a high embedding capacity of 48 bits per pixel (bpp) while maintaining minimal distortion. Specifically, the peak signal-to-noise ratio (PSNR) ranges from 34.7 to 38.12&#xa0;dB, and the structural similarity (SSIM) metric ranges from 0.9487 to 0.9643. Here, we show that IMMIH performs exceptionally well in terms of capacity, image quality, and steganalysis resistance, indicating its potential for practical applications. Detailed implementation and code can be viewed at the following link: <a href="https://github.com/ydourblne/IMMIH-framework">https://github.com/ydourblne/IMMIH-framework</a>.</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

High-capacity multi-image hiding in HDR images using invertible neural networks

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Ruo-Nan Yi,
  • Yong-Qing Huo

摘要

Deep learning has revolutionized data hiding by automating the determination of hidden positions and intensity settings through network training under specific loss functions. The large capacity of image hiding has garnered considerable attention from researchers. To address the common issues of shadow and contour distortions in multi-image hiding, we propose a novel reversible neural network framework for embedding multiple low dynamic range (LDR) images into a single high dynamic range (HDR) image. This framework, termed IMMIH, leverages the properties of invertible neural networks (INNs) to achieve fully coupled and reversible information hiding and extraction through a single end-to-end training process. By focusing on the RGB channels of HDR images, IMMIH minimizes distortion while incorporating attention mapping and edge mapping modules to improve adaptability. Furthermore, the use of discrete wavelet transform and low-frequency wavelet loss concentrates data hiding in high-frequency regions, ensuring robustness. Experimental results demonstrate that IMMIH achieves a high embedding capacity of 48 bits per pixel (bpp) while maintaining minimal distortion. Specifically, the peak signal-to-noise ratio (PSNR) ranges from 34.7 to 38.12 dB, and the structural similarity (SSIM) metric ranges from 0.9487 to 0.9643. Here, we show that IMMIH performs exceptionally well in terms of capacity, image quality, and steganalysis resistance, indicating its potential for practical applications. Detailed implementation and code can be viewed at the following link: https://github.com/ydourblne/IMMIH-framework.