Based on the hypothesis testing of the watermark bits,a model of the blind watermark extraction is presented and implemented in the wavelet domain,which can extract the watermark bit by bit. To balance the imperceptibility and robustness of the watermark,the wavelet features of the image and the context modeling method are utilized to determine the variable watermark amplitude for every high frequency wavelet coefficient. So the embedding density of the watermark is adaptive to the image features. The watermark extraction results are dramatically based on the distribution model of the image's wavelet coefficients,thereby the wavelet coefficients are modeled as a general gauss distribution,and the parameters are estimated by the maximum likelihood method. A series of experimental results show that the method is robust.