Abstract: A robust blind watermark detection technique is proposed. Through the orthogonality design between watermark signal and host data, the interference of the host data noise can be completely eliminated, which greatly improves the performance of the proposed blind watermark detector. The analysis shows that the detector outperforms other blind watermark detectors using either Neyman-Pearson criterion or Minimum Error Probability criterion. Moreover, a classification technique of watermark weights is used, which makes the performance of the detector approximate to that of watermark detectors using original host data. Experimental results verify the performance of the watermark detector under different distortions.