Rapid intelligent watermarking system for high-resolution grayscale facial images.

摘要

In this thesis, an associative memory is proposed to store the previously calculated solutions for different categories of texture using the optimization results of the whole image for few training facial images. A multi-hypothesis approach is adopted to store in the associative memory the solutions for different clustering resolutions (number of blocks clusters based on texture features), and finally select the optimal clustering resolution based on the watermarking metrics for each facial image during generalization. This approach was verified using streams of facial captures from PUT database (Kasinski et al., 2008). It was compared against a baseline system representing traditional IW methods with full optimization for all stream images. Both proposed and baseline systems are compared with respect to quality of solution produced and the computational complexity measured in fitness evaluations. The proposed approach resulted in a decrease of 95.5% in computational burden with little impact in watermarking performance for a stream of 198 facial images. The proposed framework Blockwise Multi-Resolution Clustering (BMRC) has been published in Machine Vision and Applications (Rabil et al., 2013a).;Although the stream of high dimensionality optimization problems are replaced by few training optimizations, and then recalls from an associative memory storing the training artifacts. Optimization problems with high dimensionality search space are challenging, complex, and can reach up to dimensionality of 49k variables represented using 293k bits for high-resolution facial images. In this thesis, this large dimensionality problem is decomposed into smaller problems representing image blocks which resolves convergence problems with handling the larger problem. Local watermarking metrics are used in cooperative coevolution on block level to reach the overall solution. The elitism mechanism is modified such that the blocks of higher local watermarking metrics are fetched across all candidate solutions for each position, and concatenated together to form the elite candidate solutions. This proposed approach resulted in resolving premature convergence for traditional EC methods, and thus 17% improvement on the watermarking fitness is accomplished for facial images of resolution 2048x1536. This improved fitness is achieved using few iterations implying optimization speedup. The proposed algorithm Blockwise Coevolutionary Genetic Algorithm (BCGA) has been published in Expert Systems with Applications (Rabil et al., 2013c).;The concepts and frameworks presented in this thesis can be generalized on any stream of optimization problems with large search space, where the candidate solutions consist of smaller granularity problems solutions that affect the overall solution. The challenge for applying this approach is finding the significant feature for this smaller granularity that affects the overall optimization problem. In this thesis the texture features of smaller granularity blocks represented in the candidate solutions are affecting the watermarking fitness optimization of the whole image. Also the local metrics of these smaller granularity problems are indicating the fitness produced for the larger problem.;Another proposed application for this thesis is to embed offline signature features as invisible watermark embedded in facial captures in passports to be used for individual verification during border crossing. The offline signature is captured from forms signed at borders and verified against the embedded features. The individual verification relies on one physical biometric trait represented by facial captures and another behavioral trait represented by offline signature. (Abstract shortened by UMI.).