World Applied Sciences Journal 22 (11): 1572-1580, 2013 ISSN 1818-4952 © IDOSI Publications, 2013 DOI: 10.5829/idosi.wasj.2013.22.11.2828 Corresponding Author: Mohamad Vafaei, Department of Electrical Engineering, Najafabad Branch, Islamic Azad University, Isfahan, Iran. 1572 A New Robust Blind Watermarking Method Based on Neural Networks in Wavelet Transform Domain M. Vafaei, H. Mahdavi-Nasab and H. Pourghassem Department of Electrical Engineering, Najafabad Branch, Islamic Azad University, Isfahan, Iran Abstract: In this paper, a blind watermarking method based on neural networks in discrete wavelet transform domain is proposed. Robustness and imperceptibility are main contradictory requirements of a watermark. In the proposed method, better compromises are achieved using artificial neural networks to adjust the watermark strength. A binary image is used as the watermark and embedded repetitively into the selected wavelet coefficients, which also improves the watermark robustness. Experimental results demonstrate that the proposed scheme has a simultaneous good imperceptibility and high robustness against several types of attacks, such as Gaussian and salt and pepper noise addition, cropping, mean and median filtering and JPEG compression. Key words: Blind watermarking Discrete wavelet transform Neural network INTRODUCTION add a pseudorandom pattern into host image to embed a With modern developments of information with the same pattern or by applying other statistics to technology, widespread communication networks and the watermarked image. In quantization watermarking a set digital multimedia, effective ways are needed to protect of features extracted from the host images are quantized the security of digital data. Digital watermarking is one of so that each watermark bit is represented by a quantized the proposed solutions for copyright protection in feature value. This technique improves the robustness to multimedia (such as image, sound and video), in which a JPEG compression and other typical attacks. specified hidden signal (watermark) is embedded in In [7], the host image was decomposed using 3-level digital data [1]. Robustness and imperceptibility are two DWT. The watermark was embedded into components of basic requirements of digital image watermarking that are the third decomposition layer of the DWT of an image. contradictory. Every seven non-overlap wavelet coefficients of the host Many digital watermarking algorithms have been image were grouped into a block. The differences between proposed in spatial and transform domains. The local maximum and local second maximum values were techniques in spatial domain still show relatively low modified to the watermark bit. In [11], to achieve the capacity and are not robust enough to lossy image secrecy of watermark, variable block size was used for compression and other image processing operations [2, 3]. embedding a watermark bit using different sub-bands. On the other hand, frequency domain techniques can The difference between our work and related works lie in embed more bits as watermark and are more robust to the more elaborate selection of wavelet coefficients for attacks. Transforms such as discrete Fourier transform [4], watermark embedding, the block selection process and Discrete Cosine Transform (DCT) [5] and Discrete Artificial Neural Network (ANN) inputs. Wavelet Transform (DWT) [6, 7] are generally used for In recent years, neural networks pave the way for the watermarking in the frequency domain. Watermark further development of watermarking techniques by embedding can generally be classified into two categories imitating the learning ability of brain. Neural networks are [8]: Spread Spectrum based [9] and quantization based applied either to improve watermark extraction or to watermarking [7, 10, 11]. The spread spectrum methods determine the strength of watermark [12-14]. watermark. This watermark can be detected by correlating