Genetic Algorithm-Backpropagation Network Hybrid Architecture for Grayscale Image Watermarking in DCT Domain

In this paper, Human Visual System (HVS) characteristics are modeled using a Genetic Algorithm (GA) based technique for the determination of weights in a BPN (GA/BPN) for the purpose of image watermarking. The GA based BP network is trained by 27 inference rules comprising of three input HVS features namely luminance sensitivity, edge sensitivity computed using threshold and contrast sensitivity computed using variance. The GA/BP network block wise produces a single output weighting factor which is used to embed two different watermarks - (a) a sequence of normalized random numbers and (b) a binary image, with in the host image in the transform (DCT) domain. The high computed value of PSNR parameter indicates that the signed image has good perceptible quality. The watermark is extracted from the signed image using Cox´s algorithm. The embedded and extracted watermarks are compared and SIM(X, X*) correlation parameter is computed.