Intelligent Systems for Grayscale Rapid Watermarking

Bassem Rabil, Ph.D candidate


Intelligent watermarking techniques use evolutionary computation algorithms to optimize both quality and robustness of the watermarking system. Quality fitness represents the visual impact of embedding the watermark into the host image, and robustness fitness represents the ability to extract the watermark after manipulating the watermarked image using content preserving manipulations like compression, quantization, filtering, etc. Due to the increasing demand to ensure authenticity and integrity of images which are transferred across different channels, the volume of images to be watermarked went up to tens of millions daily. Rapid watermarking systems are highly required in many domains like financial documents in banking applications, and medical images in medical imaging domain to be able to handle high volume of images to be watermarked.

Problem statement

In the case of grayscale images, watermarking metrics representing both quality and robustness are calculated for the whole image while being dependent on the embedding bands of each and every image block of the host image. This implies an optimization problem of dimension which is equal to the number of watermark bits distributed among different host image blocks.

The watermarking attacks and their priorities to be considered in the intelligent watermarking system may vary according to the nature of watermarked images. Given the high dimensional nature of the problem, it would be computationally expensive to re-optimize embedding parameters every time the attacks to be considered are changed.

Also the optimization high dimension for grayscale images implies computational complexity for each image to reach optimal embedding parameters in such high dimensional space. Optimizing a single image may take several hours to reach the optimal embedding parameters which make it infeasible to handle volume of images in range of tens of millions daily.


Multi-objective formulation for intelligent watermarking systems is proposed to be used as it fits more the nature of conflicting objectives of watermarking quality and robustness. This also gives the capability of producing multiple non-dominated solutions; this implies the flexibility of the watermarking system such that the system operator can choose among these multiple solutions according to priority of different objectives and attacks without the need for computationally expensive re-optimizations. This formulation is demonstrated in Fig. 1.

In most of application domains, the stream of grayscale images to be watermarked is of similar content. The optimization experience of the watermarking system can be utilized with subsequent images in the stream of similar content. This replaces the computationally expensive optimization of some images to memory recalls with minimal computational complexity. A novel formulation is proposed to handle the stream of images as one dynamic optimization problem with optimization performed with the first occurrence of an image and memory recalls for images of similar content. This formulation shown in Fig. 2 assumes that the environment is static during optimizing an image, and it changes whenever a new grayscale image is fed into the intelligent watermarking system. If the new image fed into the system is of similar content, this is not considered as a change in the environment, and memory recall is used rather than optimization to reach optimal embedding parameters.


Rabil, B.S., Sabourin, R. and Granger, É., Intelligent Watermarking with Multi-Objective Population Based Incremental Learning, The Sixth International Conference on Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP 2010), Darmstadt, Germany, October 15-17, 2010. (Accepted)