Intelligent watermarking of long streams of document images

Digital watermarking has numerous applications in the imaging domain, including (but not limited to) fingerprinting, authentication, tampering detection. Because of the trade-off between watermark robustness and image quality, the heuristic parameters associated with digital watermarking systems need to be optimized. A common strategy to tackle this optimization problem formulation of digital watermarking, known as intelligent watermarking (IW), is to employ evolutionary computing (EC) to optimize these parameters for each image, with a computational cost that is infeasible for practical applications. However, in industrial applications involving streams of document images, one can expect instances of problems to reappear over time. Therefore, computational cost can be saved by preserving the knowledge of previous optimization problems in a separate archive (memory) and employing that memory to speedup or even replace optimization for future similar problems. \n \nThat is the basic principle behind the research presented in this thesis. Although similarity in the image space can lead to similarity in the problem space, there is no guarantee of that and for this reason, knowledge about the image space should not be employed whatsoever. Therefore, in this research, strategies to appropriately represent, compare, store and sample from problem instances are investigated. The objective behind these strategies is to allow for a comprehensive representation of a stream of optimization problems in a way to avoid re-optimization whenever a previously seen problem provides solutions as good as those that would be obtained by reoptimization, but at a fraction of its cost. Another objective is to provide IW systems with a predictive capability which allows replacing costly fitness evaluations with cheaper regression models whenever re-optimization cannot be avoided. \n \nTo this end, IW of streams of document images is first formulated as the problem of optimizing a stream of recurring problems and a Dynamic Particle Swarm Optimization (DPSO) technique is proposed to tackle this problem. This technique is based on a two-tiered memory of static solutions. Memory solutions are re-evaluated for every new image and then, the re-evaluated fitness distribution is compared with stored fitness distribution as a mean of measuring the similarity between both problem instances (change detection). In simulations involving homogeneous streams of bi-tonal document images, the proposed approach resulted in a decrease of 95% in computational burden with little impact in watermarking performace. Optimization cost was severely decreased by replacing re-optimizations with recall to previously seen solutions. \n \nAfter that, the problem of representing the stream of optimization problems in a compact manner is addressed. With that, new optimization concepts can be incorporated into previously learned concepts in an incremental fashion. The proposed strategy to tackle this problem is based on Gaussian Mixture Models (GMM) representation, trained with parameter and fitness data of all intermediate (candidate) solutions of a given problem instance. GMM sampling replaces selection of individual memory solutions during change detection. Simulation results demonstrate that such memory of GMMs is more adaptive and can thus, better tackle the optimization of embedding parameters for heterogeneous streams of document images when compared to the approach based on memory of static solutions. \n \nFinally, the knowledge provided by the memory of GMMs is employed as a manner of decreasing the computational cost of re-optimization. To this end, GMM is employed in regression mode during re-optimization, replacing part of the costly fitness evaluations in a strategy known as surrogate-based optimization. Optimization is split in two levels, where the first one relies primarily on regression while the second one relies primarily on exact fitness values and provide a safeguard to the whole system. Simulation results demonstrate that the use of surrogates allows for better adaptation in situations involving significant variations in problem representation as when the set of attacks employed in the fitness function changes. \n \nIn general lines, the intelligent watermarking system proposed in this thesis is well adapted for the optimization of streams of recurring optimization problems. The quality of the resulting solutions for both, homogeneous and heterogeneous image streams is comparable to that obtained through full optimization but for a fraction of its computational cost. More specifically, the number of fitness evaluations is 97% smaller than that of full optimization for homogeneous streams and 95% for highly heterogeneous streams of document images. The proposed method is general and can be easily adapted to other applications involving streams of recurring problems.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.