Attention Driven Solutions for Robust Digital Watermarking Within Media

As digital technologies have dramatically expanded within the last decade, content recognition now plays a major role within the control of media. Of the current recent systems available, digital watermarking provides a robust maintainable solution to enhance media security. The two main properties of digital watermarking, imperceptibility and robustness, are complimentary to each other but by employing visual attention based mechanisms within the watermarking framework, highly robust watermarking solutions are obtainable while also maintaining high media quality. This thesis firstly provides suitable bottom-up saliency models for raw image and video. The image and video saliency algorithms are estimated directly from within the wavelet domain for enhanced compatibility with the watermarking framework. By combining colour, orientation and intensity contrasts for the image model and globally compensated object motion in the video model, novel wavelet-based visual saliency algorithms are provided. The work extends these saliency models into a unique visual attention-based watermarking scheme by increasing the watermark weighting parameter within visually uninteresting regions. An increased watermark robustness, up to 40%, against various filtering attacks, JPEG2000 and H.264/AVC compression is obtained while maintaining the media quality, verified by various objective and subjective evaluation tools. As most video sequences are stored in an encoded format, this thesis studies watermarking schemes within the compressed domain. Firstly, the work provides a compressed domain saliency model formulated directly within the HEVC codec, utilizing various coding decisions such as block partition size, residual magnitude, intra frame angular prediction mode and motion vector difference magnitude. Large computational savings, of 50% or greater, are obtained compared with existing methodologies, as the saliency maps are generated from partially decoded bitstreams. Finally, the saliency maps formulated within the compressed HEVC domain are studied within the watermarking framework. A joint encoder and a frame domain watermarking scheme are both proposed by embedding data into the quantised transform residual data or wavelet coefficients, respectively, which exhibit low visual salience.