Abstract: The detection errors undermine the credibility of a watermarking system. An optimum image blind watermark detector is designed in this paper in discrete wavelet domain, which adopts a dual channel detection using bispectrum detection and energy detection. Even with small signal noise ration, the detector can still get a high detection probability with enough bispectum information. When the bispectrum is small, this dual channel detection will become a likelihood detection system, and it is still an optimum detection system. Experimental results demonstrate that the proposed optimum watermarking detector is robust with respect to attacks produced by the popular watermark test software - Stirmark, such as cropping, JPEG compression, rotation, scaling, random geometrid distortion. The experimental results also show that this optimum image blind watermark detector has superior advantages than the exist watermark detectors.