In this paper

a blind meaningful watermarking algorithm is proposed. The host image is first transformed into wavelet domain where the feature set is extracted. Through subspace decomposition of the feature set

the watermark signal sets are generated which holds orthogonality between each two watermark signals and between the watermark signals and the feature set. For the orthogonality between the watermark signals and the feature set

the meaningful watermark can be extracted without resorting to the host image. And for the orthogonality between each two watermark signals

the characters embedded can be retrieved robustly using matched filtering. Because the watermark signal sets are image content dependent

the algorithm security can be improved. In order to measure the reliability of the extracted characters

the concept of credibility threshold is proposed by which the wrong characters can be correctly pointed out. Therefore the credibility of the whole message can also be evaluated

which is another contribution of this paper. In the experiment

we test the algorithm performance under JPEG compression and Additive White Gaussian Noise(AWGN). The results show that the robustness and imperceptibility of the watermark are satisfactory. The credibility threshold also proves its effectiveness.