Digital watermarking has emerged as a potentially effective tool for multimedia copyright protection, authentication and tamper proofing. Various approaches to watermarking have been proposed which exploit image properties and image processing techniques in different domains. The majority of the algorithms developed to date are concentrated in spatial, frequency, and wavelet domains. In recent years, image watermarking based on image moments has drawn more interest. This is due to the fact that image moments have been shown to be robust with respect to image variations such as in scale or position. Recently, H. S. Kim and H. K. Lee [24] proposed a new image watermarking technique using Zernike moments (ZM). ZM are rotation invariant and they can be made to be both scale and translation invariant by properly normalizing the image. In addition, it also has very good resistance to noise in the image. The fundamental problem with this the existing moments based watermarking techniques is that the watermark synchronization between source and receiver is highly dependent upon the same image centroid location is detected at both the source and receiver. As a result, they are vulnerable against geometric attacks involves image cropping or padding.

In the research, a new blind image watermarking scheme using a novel pattern recognition framework is presented. In both watermark embedding and extraction, we first utilize an affine invariant point detector to determine an invariant point as the center and use the scale associated with this point to determine the region to embed/detect the watermark. When embedding the watermark, a self-similar sequence serves as a quantization function to modulate the amplitude of the image moments with a pseudo random sequence. The watermark detection consists of a weighted correlation of the original pseudo random sequence with a binary sequence extracted by demodulating the amplitude function by searching through the same self-similar quantization function. The performances of two types of image moments (ZM and Radial Tchebichef Moments (RTM)) in this framework are studied. Experiments conducted using Stirmark4 were demonstrate that this proposed framework overcomes the limitation of existing moments-based approaches in dealing with geometric attacks involves image cropping/padding operations. The framework is then extended to video watermarking realm, where two video watermarking implementations that utilize RTM and Hu’s moment invariants (MI) respectively are discussed.