A robust RST invariant image watermarking method based on locally detected features

This work presents a robust rotation, scaling and translation (RST) invariant image watermarking method based on locally detected features. First, the feature points are detected in the original image by using the Harris corner detection. Then, the Delaunay tessellation is created on these feature points. Watermark data are embedded and detected in each of the Delaunay triangles. The Harris corner detection is the key point in this scheme. Traditionally, the global threshold is used for the Harris corner detection. As a result, the points are concentrated on areas with rich texture information and the contrast change within the image. In this way, the locations of points are seriously affected by attacks, such as, geometrical attacks. It makes the watermark detection difficult or even impossible. In order to generate more robust feature points, we use a local threshold feature detection method, in which local thresholds are used for different parts of the target image. The feature points are almost uniformly distributed on the image. The experimental results show that this method makes feature points very robust after geometrical attacks, JPEG compression and noise addition.