MLP Neural Network to improve Digital Watermark detection in gray scale images

Abstract

Image watermarking has today a growing success in the community of image processing. Many methods were already proposed making it possible to obtain increasingly more powerful algorithms in spatial and frequency domain. Most of the spatial watermarking schemes are based on the image decomposition into a grid of blocks, in order to insert a sequence of bits (message). To reach good performances, content-based watermarking schemes aim to use feature points to link the mark with the content of the image [8]. The detection step of all these methods perform a thresholding operation on the correlation function, computed with respect to the block to be processed and the mark or the estimated mark using Wiener filtering method [14]. It’s a hard task to fix the threshold value to be used in this step and any improvement here can enhance performances of the global scheme. In this paper we look for a suitable alternative to perform this task easily and to improve the detection step by using artificial neural networks. In fact, a training phase is performed using a MLP neural network that can be feed by an image block and gives a float value as an output that we can use to take a decision about the presence of the mark. Although, the training phase is time consuming, it’s performed separately. This method gives good results even when mark estimation using Wiener filtering is not used.