In this paper, an optimal estimation of the embedding strength for the digital image watermark is discussed. Under the assumptions that the original image data are an i.i.d. random sequence from a Guassian distribution, it's shown that this embedding strength should be estimated by observing the averages of the watermarked image and briefly calculating between them. The approach is based on maximum-likelihood estimation of the watermark embedding strength. The experimental results show that this method is effective for watermark extraction. Moreover, according to the addition, multiplication and fusion embedding rules, their performances are compared and analyzed using PSNR. The conclusions obtained by experiments are useful to future research.