Digital image watermarking, as an extension of traditional steganography, refers to the process of hiding certain messages into cover images. The transport image, called marked-image or stego-image, conveys the hidden messages while appears visibly similar to the cover-image. Therefore, image watermarking enables various applications such as copyright protection and covert communication. In a watermarking scheme, fidelity, capacity and robustness are considered as crucial factors, where fidelity measures the similarity between the cover- and marked-images, capacity measures the maximum amount of watermark that can be embedded, and robustness concerns the watermark extraction under attacks on the marked-image. Watermarking techniques are often trade-offs between these factors; for example, a high capacity usually implies more modification on the cover-images and thus lowers the fidelity, and the robustness often applies redundancy and lowers capacity. Traditional image watermarking schemes place the watermark on the trivial portions of cover images to enable the invisibility; however, the hiding can be easily revealed by statistical analysis. Hence, during recent years, researchers have proposed different image watermarking schemes aiming at improvements from various perspectives, such as embedding the watermark into the frequency spectrum for high fidelity and high security, extending the capacity via iterative embedding, enhancing the undetectability by maintaining the image statistics and improving the robustness applying statistical features like image histogram. But the adaptation to varying, flexible or multi-purposed situations remains a challenge in existing watermarking methods due to the randomness of the cover image contents. In addition, fewer attempts have been reported to level the trade-off when two or more controversial watermarking factors are required. Moreover, although computational intelligence has grown rapidly in the past decade, applying its adaptation ability in image watermarking remains a gap. In this dissertation, some adaptive image watermarking schemes are presented. First, to achieve content adaptation on the spatial domain, a novel salient region detection model is presented to automatically segment the cover images into regions-of-interests (ROIs) and region-of-noninterests (RONI). The ROIs containing the most representative information are kept intact during the embedding and the RONI is collated for watermarking. Second, an intelligent image watermarking scheme based on the ROI detection is presented. A novel reversible watermarking algorithm that achieves high capacity and low distortion is firstly introduced. It is then followed by partitioning algorithms to bridge the gap between ROIs based image watermarking schemes and watermarking embeddings on frequency domain. Partition ranking schemes based on entropy as well as swarm intelligence are proposed, not only to optimize the overall watermark embedding, but also to provide flexibility that the watermarking purpose can be determined by the end user. Third, to conquer the robustness issue, a robust image watermarking scheme based on the ROIs detection is presented. With a robust watermarking algorithm based on contrast modulation, it matches the segmented ROIs between the marked-image and the distorted image to rectify the attacks. Finally, an image watermarking system using deep learning is introduced, where the rules of watermark embedding and extraction are learned and generalized in an unsupervised manner, which is fully adaptive to image contents and features.
|
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
