Abstract
For the shortcoming that transform domain watermarking cannot resist image rotation attack, this study draws LPT (Log-Polar-Transform) theory into the watermarking field, then puts forward a new improved algorithm based on LPT and DFT. Before embedding the watermark, the carrier image undergoes LPT and DFT. Because LPT and DFT theories respectively have good scale invariability and synchronicity when image rotation happens, they provide an environment capable of resisting rotation. The watermark information spectrum is then spread with a pseudorandom sequence and embedded into the low-frequency band of the carrier. Experimental results prove that the algorithm manifests good robustness when attacked by image rotation and other common image processing methods.
Introduction
Digital watermarking is essential for copyright protection, but it has a notorious "Achilles' heel": geometric attacks. While many transform-domain methods (like those using DCT or DWT) are robust against noise or compression, they often fail the moment an image is rotated. When an image rotates, the synchronization between the embedded watermark and the carrier pixels is lost, making the watermark appear as random noise to standard detection algorithms. This research aims to create a "rotation-resistant" environment where the watermark's mathematical signature remains stable regardless of orientation.
Methodology
The proposed algorithm leverages the unique properties of the Log-Polar domain to decouple rotation from the watermark signal. The pipeline follows several key stages:
- Log-Polar Transform (LPT): The carrier image is converted from Cartesian coordinates $(x, y)$ to Log-Polar coordinates $(r, heta)$. LPT converts physical rotation and scaling into simple vertical and horizontal shifts in the transform domain.
- Discrete Fourier Transform (DFT): The LPT-transformed image is processed using a 2D DFT to provide frequency stability.
- Spread Spectrum Embedding: The watermark (a $64 \times 64$ binary image) is spread using a pseudorandom sequence and embedded into the low-frequency bands of the DFT coefficients.
- Perceptual Tuning: A "brightness shield" based on the Human Visual System (HVS) adjusts the embedding strength ($\alpha$), ensuring the watermark remains invisible to the human eye.
Experimental Results
The algorithm was evaluated using a standard $512 \times 512$ grayscale "Couple" image as the carrier and a binary "LNTU" text image as the watermark. Success was measured using the Normalized Correlation (NC) value, where 1.0 represents a perfect match.
Rotation Resistance
The core strength of the LPT+DFT approach is its resilience to geometric distortion. The algorithm maintained high NC values across various rotation angles:
- 30° Rotation: NC = 0.9033
- 45° Rotation: NC = 0.8758
- 90° Rotation: NC = 0.9671
General Robustness
Beyond rotation, the watermark proved highly resilient to non-geometric attacks due to its placement in low-frequency bands:
- White Noise: NC = 0.9871
- JPEG Compression: NC = 0.9835
- Pepper Noise: NC = 0.9812
Limitations and Future Work
While the algorithm effectively solves the rotation problem, it remains vulnerable to intensity-based attacks. Specifically, histogram equalization (NC = 0.5166) and contrast changes (NC = 0.3945) significantly degrade the frequency-domain signature. Additionally, while LPT theoretically handles scaling, experimental results for resizing attacks showed visible distortion in extracted watermarks. Future research will focus on improving robustness against contrast manipulation and extreme geometric scaling.
Summary updated