Geometrical Invariant Image Blind Watermarking Techniques Based On Feature Points

Nowadays, through Internet and CD-ROM, the digital copies with the same quality as original data can be got much more easily and quickly. But the consequential problems such as copyright piracy, illegal use and copyright dissension have increasingly become a serious social issue.As a potential effective solution to the aforementioned problems, digital watermarking technology has received extensive attention. Because of the universality and important role of the digital image in the multimedia data, and it representative research process, the digital image watermarking is mainly taken as the research object.As a means of copyright protection, the digital watermarking will suffer various attacks, thus the watermarking for copyright protection should have the robustness. However, the robustness of the existing watermarking schemes against the geometrical attacks is still not enough, especially in the blind watermarking applications. The second generation watermarking, namely the content-based localized watermarking scheme, is promising to resist to the geometrical attacks. The basic idea of the content-based localized watermarking scheme is that, the robust feature point of the image is used to locate the watermark embedding and the watermark is independently embedded into the corresponding patch to each feature point so that the watermark can still be located and detected by the feature points of the part image, and thus can resist to the cropping attacks effectively. The content-based localized watermarking, also called the feature-based localized watermarking can logically divided into three parts: the detection of feature points, the construction of the elementary patches used to embed and detect the watermark, and the watermarking algorithm of elementary patches. The robustness of the whole watermarking scheme depends on the robustness of the feature points, the robustness of the constructed patches based on the feature point, and the robustness of the watermarking algorithm of the patch.Most of the current existing feature-based watermarking algorithms directly use the feature from the pattern recognition fields to synchronize the watermark. However, the requirements in the watermarking applications and the pattern recognition fields are not identical and the feature used for the pattern recognition fields is not necessarily suit for the watermarking applications. Thus, based on the mean luminance of the disk, a rotation and scale invariant feature (shortly named DRSIF)extraction algorithm used for the watermarking applications is proposed.For each point of the image, centered at which, firstly a series of concentric rings are generated and the absolute difference between the mean luminance of the inner ring and the mean luminance of the outer ring is computed for each ring when the inner radiuses of the rings come from a discrete geometric progression and the ratio of the outer radius to the inner radius of each ring is fixed, then the maximum of absolute mean luminance difference sequence is taken as the response value of the observed point, and the inner radius of the ring which attains the maximum of absolute mean luminance difference sequence is taken as the characteristic radius of the observed point. Finally the points whose response value attains the local maximum are selected as DRSIF points. In comparison to the two traditional scale invariant feature points, the time complexity of the DRSIF extractor is analyzed and the robustness of the DRSIF pints are verified. The experimental results show that the DRSIF points are robust against common attacks in the watermarking applications and are suit for the watermarking applications.The feature points are utilized to divide the image into the elementary patches used to embed and detect the watermark before the watermark is embedded and detected. However, the geometrical invariability of the feature points can not necessarily ensure that the constructed patches have the same geometrical invariability. The geometrical invariability of the patch also depends on the specific division, namely the way to construct the patch. How to construct the circular patches used to embed and detect the watermark is respectively elaborated. A partitioning strategy is proposed to improve the hit ratio of the embedding patches during the watermarking detection so that the speed to find the watermarked patch is greatly increased. Through experiments, the robustness of the elementary patch used to embed the watermark is verified and the proposed partitioning strategy during the watermarking detection is proved to be effective.Based on the aforementioned research achievements, two complete DRSIF-based localized watermarking schemes are proposed: one is DRSIF-based localized watermarking scheme in the spatial domain, and the other is localized watermarking scheme based on DRSIF points and invariant domain. In the former scheme, the moment of the patch is used to compute the characteristic angle of the embedding patch, which can reflect the rotation distortions suffered by the image, so that the rotation invariant of the patch is attained. In the latter scheme, the watermark is independently embedded into the rotation and scale (shortly named RS)invariant domain of the embedding patch. By combining the RS invariant domain with the idea of the localized watermarking, the watermark can resist to not only the whole geometrical attacks, but also the cropping attacks and such local geometrical attacks. Through experiments, the two watermarking schemes are proved to be robust. In comparison to the other algorithms, the performances of the proposed schemes perform better. The proposed two localized watermarking schemes are further compared. The robustness of the localized watermarking scheme based on DRSIF points and invariant domain is better than the other scheme because the latter requires more accurate repeatability of the feature points. In addition, the watermark volume of the localized watermarking scheme based on DRSIF points is less than the other. Thus, In the concrete application, the watermark embedding algorithms and detection algorithms are carefully selected by considering the requirement of the robustness and the watermark volume.