A New Algorithm for Blind Image Watermark Detection
Bo Xiao · Chinese Journal of Computers · 2001
Digital watermarking is a key technique for protecting intellectual property of digital media. Due to the ability to detect watermark without the original image, blind watermarking is very useful if there are too many images to be authenticated. Since Cox et al. proposed a DCT-based spread spectrum approach to hide watermark, a lot of watermarking schemes in the DCT domain have been presented. Barni improved Cox's algorithm and made it a blind watermarking scheme by embedding the signature in the fixed position. But the watermark algorithm in is based on the calculation of the correlation coefficient between the image and the watermark in the DCT domain. In this paper, we began our analysis by posing the difference on mathematical models between private watermark and blind watermark detection. The watermark detection in is not detecting weak signal in noise, but comparing watermark signal and its estimation. Based on this analysis, we point out that the linear correlation detector, which has been somewhat taken for granted in the previous literature in the private watermarking algorithms, would be optimal in blind watermark only if the host signal followed Gaussian distribution. After reviewing some statistical models which have been proposed to better characterize the DCT coefficients of images, we find that the popular Gaussian distribution is not accurate enough to model the peaky, heavy-tailed marginal distribution of DCT coefficients. So we devise and implement a new blind watermark detector--sign correlation detector based on the Laplacian distribution model of AC DCT coefficient. Computing result of asymptotic relative efficiency demonstrates that efficiency of sign correlation detector is much greater than that of linear correlation. A series of experiments also show that it is more robust than linear correlation detector. After compressing the watermarked image using JPEG standard with 75% quality, or filtering the watermarked image three times using median filter, or adding white noise to watermarked image with very low SNR, the linear correlation detector can not detect the watermark correctly, while the sign correlation detector proposed in this paper still outputs satisfactorily.