Research On Digital Watermark Based On Neural Networks

Embedding and extracting are the critical steps for watermarking algorithm. Through studying the watermarking on digital images, how to improve the stability of watermarking by using neural net has been analyzed in this subject. Two aspects are discussed here, namely, how to improve the accuracy of monitoring by using neural net to detect the watermarking and how to improve the stability of embedded watermarking.People are the subject of image quality evaluation, almost all the image processing techniques are related to Human Visual System (HVS). HVS is a non-linear system, the complexity, vein and luminance of image are important for the invisibility of watermarking embedding. To solve the contradiction between the invisibility and stability of digital watermarking, the characteristics of HVS must be considered and the energy of watermarking signal should be allocated reasonably, so as to improve the intensity of local watermarking embedding as far as possible without affecting perception.Neural net has high ability of non-linear glob fitting and fault-tolerance, and great function of self-adaption and self-learning. Its essence reflects a way transforming input to output with non-linear process mode, and can solve some problems which traditional algorithm cannot solve. Neural net can approach non-linear function infinitely; it has similar nature with the inverse process of picture processing and can improve the effect of watermarking extraction and watermarking detection effectively.This subject presents the technology of using neural net to extract watermarking based on the research of the correlation detection for existing spatial domain and neighborhood. Firstly, embed a certain amount of redundant information in watermarking signals and neural net, train the neural net in training mode created by certain redundant information when extracting, embed the correlation of pixel value and neighborhood difference value of information position into the neural net memory, and then obtain the actual judgment threshold of watermarking signal with neural net dynamically. The experiment results show that it has good robustness in geometric processing, like image rotation.Wavelet has good multi-resolution analysis and local-time frequency characteristics, energy concentration characteristic both in frequency domain and spatial domain, and has high spatial correlation between sub-image and pixel in the corresponding spatial position. The judgment threshold of watermarking signal with embedded redundant information in wavelet domain are obtained dynamically. The experiment results show that the signal extraction of approximation sub-band embedded wavelet has good robustness, and the sensibility for the operation like rotation is better than before.In the comprehensive consideration of HVS and neural net, a self-adaption watermarking algorithm has been presented in this paper. It collected the characteristics of frequency, vein, comentropy and luminance masking after the image is divided into blocks and the maximum embedment strength confirmed from experiments in the way of neural net training mode. The experiment results show that neural net can obtain great embedment strength, ensures the invisibility of watermarking image and has good adaption.In this paper, all the experimentation of processing algorithm has been provided, and performance evaluation has been proceeded. The test results show that the algorithm has definite stability and invisibility.