Robust Watermarking Approach Based on Coefficient Comparison of Detail Sub-bands of DWT

Qiudong Sun, Yufeng Shao, Liandong Wang, Lin Gui, Wenying YanAdvances 2014 · DOI

Abstract

This paper presented a novel watermarking method to protect copyrights for digital media. Firstly, a visual model is employed to calculate the block based just noticeable difference (JND) to control the embedding intensity to improve the transparency and robustness of watermarking. Then, two neighboring image blocks are selected each time in Hilbert scanning order, and transformed by discrete wavelet transform (DWT) respectively. Lastly, according to the JND and the watermark bit, the values of two corresponding detail sub-bands coefficients in DWT domain is adjusted to embed the watermark. In this algorithm, a couple of coefficients in two corresponding detail sub-bands can embed a bit of watermark sequence. This embedding method guarantees hiding enough copies of watermark into the host image and improves its performance. The experimental results show that the embedded watermark is invisible, and the algorithm is robust to general image processing operations. The conclusion is that the algorithm is effective and practical.

Introduction

Digital watermarking faces a fundamental "tug-of-war" between two requirements: transparency (the watermark should not degrade visual quality) and robustness (the watermark must survive common image processing "attacks"). While many methods exist in the Discrete Wavelet Transform (DWT) domain, few exploit the spatial relationships between corresponding detail sub-bands of neighboring image blocks.

This paper proposes a robust watermarking technique that embeds data by comparing wavelet coefficients of neighboring image blocks selected via a Hilbert scanning path. By using "Just Noticeable Difference" (JND) thresholds to govern embedding intensity, the system achieves a high degree of transparency while remaining resilient against attacks like JPEG compression, cropping, and noise.

Methodology

The proposed algorithm follows a multi-step pipeline designed to balance visibility and durability:

  1. Visual Perceptual Modeling: A mathematical model calculates the Just Noticeable Difference (JND) for each image block. This value acts as a "noise budget," determining how much a block's coefficients can be altered before changes become perceptible.
  2. Hilbert Scanning: Instead of a standard raster scan, the algorithm uses a Hilbert curve. This fractal-like path ensures that "neighboring" blocks are spatially close, typically sharing similar textures and lighting, which stabilizes the comparison.
  3. DWT Transformation: Each block is decomposed into frequency sub-bands (Horizontal, Vertical, and Diagonal details).
  4. Coefficient Comparison Embedding: To embed a bit, the algorithm adjusts the values of corresponding coefficients between two neighboring blocks. For example, to embed a "1," the coefficient in the first block is made significantly larger than the one in the second block, within JND limits.
Relationship between PSNR, JND range, and block size.

As shown in the figure above, the choice of block size and JND range directly impacts the Peak Signal-to-Noise Ratio (PSNR), which is a proxy for transparency. Smaller block sizes (e.g., 2x2) generally result in lower PSNR for the same JND range compared to larger blocks (e.g., 8x8).

Experimental Setup and Datasets

The authors evaluated their approach using standard grayscale test images to ensure comparability with existing literature.

  • Host Images: Classic 512×512 pixel images such as "Elaine" and "Lena".
  • Watermark: A 32×32 binary icon or logo.
  • Evaluation Metrics: PSNR was used to measure transparency, while the Normalized Correlation (NC) coefficient measured watermark recovery accuracy after various attacks.
Standard host image "Elaine" used for testing.

Key Results and Discussion

The experimental results demonstrate that the algorithm successfully navigates the transparency-robustness trade-off:

  • High Transparency: Watermarked images achieved PSNR values between 36.6dB and 36.9dB. Visual differences were negligible, only becoming apparent under 30x magnification.
  • Attack Resilience: The system showed near-perfect recovery (NC = 1.0) against brightness/contrast enhancements, histogram equalization, and salt & pepper noise.
  • Robustness to Cropping: Due to redundant encoding (repeating the watermark throughout the image), the watermark was recovered with an NC of 0.9981 even after losing a significant central portion of the image.
  • Compression Survival: The watermark remained highly legible even after 90% quality JPEG compression.

The core innovation lies in Dynamic Intensity Control. By being "content-aware," the algorithm embeds data more aggressively in highly textured areas where changes are harder to see, and more gently in smooth areas. Basing the watermark on relative differences between blocks also makes it naturally resistant to global adjustments like brightness changes.

Conclusion

The proposed DWT-based watermarking algorithm effectively utilizes JND modeling and Hilbert scanning to provide a robust and transparent copyright protection mechanism. While currently focused on grayscale images, future work could extend this logic to color spaces (RGB or YCbCr) and further investigate performance against complex geometric distortions like rotation or shearing. The high theoretical capacity and redundancy make it a practical choice for digital media security.

Summary updated

Abstract

This paper presented a novel watermarking method to protect copyrights for digital media. Firstly, a visual model is employed to calculate the block based just noticeable difference (JND) to control the embedding intensity to improve the transparency and robustness of watermarking. Then, two neighboring image blocks are selected each time in Hilbert scanning order, and transformed by discrete wavelet transform (DWT) respectively. Lastly, according to the JND and the watermark bit, the values of two corresponding detail sub-bands coefficients in DWT domain is adjusted to embed the watermark. In this algorithm, a couple of coefficients in two corresponding detail sub-bands can embed a bit of watermark sequence. This embedding method guarantees hiding enough copies of watermark into the host image and improves its performance. The experimental results show that the embedded watermark is invisible, and the algorithm is robust to general image processing operations. The conclusion is that the algorithm is effective and practical.
BibTeX
@inproceedings{sun2014robust,
  author = {Qiudong Sun and Yufeng Shao and Liandong Wang and Lin Gui and Wenying Yan},
  title = {{Robust Watermarking Approach Based on Coefficient Comparison of Detail Sub-bands of DWT}},
  booktitle = {Advances},
  year = {2014},
  doi = {10.2991/icfcce-14.2014.31},
}
Figure