Robust Image Watermarking Based On Local Features

With the rapid development of digital technologies and Internet, the scope and depth of information dissemination have been greatly expanded. Moreover, the efficiency and accuracy of information expression have been significantly improved. On the other hand, digital media can be easily copied, manipulated and redistributed without any permission from the copyright owner, which is potentially capable of incurring considerable financial loss to the media producers and content providers. Therefore, how to protect the copyright of digital media in open networks has been a challenging issue and also has become a hot topic in both academic community and business circles. Under this circumstance, the concept of digital watermarking came up while trying to solve the problems related to the management of intellectual property of digital media.After decade development, digital watermarking is still in its infancy and remains a great deal of open problems demanding prompt solution in theory and application. Under the framework of the second generation watermarking, this paper takes still images as main research object and presents several geometrically resistant watermarking algorithms by incorporating the advantages of the local invariant feature regions and the embedding/detection strategies seamlessly. The main achievements of this paper are summarized as follows.(1) In feature-based image watermarking, the stability and distribution of local feature regions will direct affect the robustness of the watermarking system. Taking this situation into account, this paper proposes a novel selection criterion for watermarking references. After choosing the feature points in the middle-scale band, this criterion utilizes graph theory clustering to mine the latent distant information betweeen feature points and groups these feature points according the distance constraint. With regarding to the same cluster, the points whose strength is the largest are used to form the feature regions. Experimental results show that this criterion is not only suitable for different types of feature point detector and image, but also stable under various watermark attacks.(2) Nowadays, the existing second generation watermarking methods are still vulnerable to some particular geometric distortions, such as local random bending, shearing, and aspect ratio changes. With regarding to this, a geometrically robust image watermarking approach is developed via affine covariant regions extracted by Harris-Affine detector. These regions have good spatial localization, orientation selection, distinctiveness, and more importantly, their covariance records not only the geometric 2D transformations but also nonrigid deformations. From the circular watermark, the elliptical watermark pattern can be obtained by the affine transform- ation according to the shape of the feature regions. The elliptical watermark is embedded additively in the spatial domain. The experimental results illustrate that the proposed scheme can resist kinds of geometric attacks.(3) To further improve the robustness against geometric attacks, an image watermarking method is proposed based on normalized affine covariant regions. By analyzing the relationship of the identical regions in both the original image and the affine-transformed version, local normalization technique is first used to reduce the affine ambiguity to a rotational one. Then, orientation alignment is applied to remove the rotation effect. In this way, geometrically invariant local feature regions can be achieved. Compared with the original method based on affine covariant regions, this new approach evidently improves the performance in terms of robustness.(4) The previous feature-based image watermarking methods are difficult to obtain the relatively higher watermark detection ratio under common image processing as well as geometric attacks. To this end, a robust image watermarking by using local invariant features is proposed. Local invariant feature regions are first constructed with Scale Invariant Feature Transform (SIFT), and then watermarking embedding and detection are conducted in DFT domain of these invariant regions. Extensive experimental results confirm that this algorithm enhances not only the overall robustness bust also the detection accuracy.(5) To deal with the shift problems and interpolation error problems caused by common image processing and geometric attacks, a local Tchebichef moments based robust image watermarking is developed. The Tchebichef moments are employed to describe the global characteristics of the local invariant regions which are relatively independent to the location deviation and the numerical error in pixel values. Here, Tchebichef moments have not only insensitivity to noise but also better feature representation capability and reconstruction accuracy. Experiments using a subset of USC-SIPI image database demonstrate the newly developed algorithm outperforms some representative methods consistently in terms of watermark capacity, impercepti- bility, and robustness.On the basis of solving the conflict between the stability and the spatial distribution of local feature regions, the construction of geometrically invariant local feature regions is combined with the design of watermark embedding/detection strategy. The proposed algorithms can effectively resist common image processing, geometric attacks, and even combined attacks. In generally, these research results enrich the theories and applications of image watermarking robust against geometric attacks.