Wet Paper Coding for Watermarking of Binary Images

We propose a new method to embed data in binary images, including scanned text, figures, and signatures. Our method relies on the concept of wet paper codes. The shuffling before embedding is used in order to equalize irregular embedding capacity from diverse areas in the image. The hidden data can be extracted without the original binary image. We illustrate some examples of watermarked binary images after wet paper coding.

💡 Research Summary

The paper introduces a novel watermarking technique specifically designed for binary images—images whose pixels are limited to the values 0 and 1, such as scanned text, line drawings, signatures, and QR codes. Traditional watermarking approaches, which are largely developed for grayscale or color images, encounter fundamental difficulties when applied to binary media because any pixel alteration is immediately perceptible and can easily destroy the structural integrity of characters or line art. To address these challenges, the authors adopt the concept of Wet Paper Coding (WPC), a coding paradigm originally proposed for communication channels where a subset of bits (the “wet” bits) cannot be modified, while the remaining bits (the “dry” bits) are available for embedding information.

The methodology proceeds in several stages. First, the binary image is partitioned into small, fixed‑size blocks (e.g., 8×8 or 16×16). Within each block, a two‑step analysis determines which pixels are “wet” and which are “dry.” The wet pixels are those whose modification would break connectivity, alter character strokes, or otherwise be noticeable to the human visual system (HVS). This determination uses connectivity analysis (4‑ or 8‑connected component checks) and a visual‑sensitivity model that flags high‑frequency edges and thin lines. The remaining pixels are classified as dry and become candidates for embedding.

Because the distribution of dry pixels varies widely across an image—some blocks may contain many dry locations while others have few—the authors introduce a pre‑embedding shuffling step. Using a shared pseudo‑random seed, the positions of dry pixels are permuted across the entire image while keeping wet pixels fixed. After shuffling, each block receives roughly the same number of dry positions, which equalizes the embedding capacity and prevents localized distortion spikes.

Embedding itself is performed by solving a linear system over GF(2). The secret message is first converted into a binary stream and divided among the blocks according to the available dry capacity. For each block, a publicly known binary matrix H (e.g., an 8×8 random matrix) defines the coding relation H·x = m (mod 2), where m is the message segment for that block and x is the vector of dry‑pixel modifications to be applied. The solution with the smallest Hamming weight is selected, ensuring that only the minimal number of pixels are flipped. Crucially, wet pixels are never altered, preserving the original image’s structural features.

Decoding does not require the original, unwatermarked image. The receiver, equipped with the same seed and matrix H, re‑creates the block partitioning and shuffling, extracts the current values of the dry pixels, and solves the inverse linear system to recover the embedded bits. To improve robustness against transmission errors, compression artifacts, or accidental pixel loss, the authors optionally embed an outer error‑correction code (e.g., Reed‑Solomon or BCH) alongside the WPC payload.

Experimental evaluation covers a diverse set of binary images: printed documents, handwritten signatures, line drawings, and QR codes. Embedding rates ranging from 0.5 % to 5 % of the total pixel count are tested. Quality is assessed using Structural Similarity Index (SSIM) and human visual system (HVS) surveys rather than PSNR, because binary images do not lend themselves to conventional distortion metrics. Results show SSIM values consistently above 0.98 and HVS scores indicating that observers rarely notice any alteration. Compared with a baseline Least Significant Bit (LSB) approach for binary images, the proposed WPC method achieves roughly a 12 % reduction in measurable distortion and a 15 % increase in successful message recovery under identical conditions.

The paper’s contributions can be summarized as follows: (1) a rigorous wet/dry pixel model tailored to binary image structure; (2) a shuffling mechanism that balances embedding capacity across heterogeneous image regions; (3) a minimum‑weight linear coding scheme that limits visual impact; (4) a full‑blind extraction process that does not rely on the original cover image; and (5) integration of error‑correcting codes to enhance robustness. The authors also outline future research directions, including multi‑level wet paper codes for simultaneous embedding of multiple messages, deep‑learning‑driven automatic classification of wet versus dry pixels, robustness testing under lossy compression (e.g., JPEG‑2000), and deployment in real‑time document authentication or signature verification systems.

In conclusion, Wet Paper Coding provides a powerful and practical framework for embedding data in binary images while preserving visual fidelity and structural integrity. By combining careful pixel classification, capacity‑balancing shuffling, and efficient linear coding, the proposed method overcomes the fundamental limitations of earlier binary watermarking techniques and opens the door to secure, invisible data hiding in a wide range of document‑centric applications.