Content analysis technologies give more and more emphasis on multimedia semantics. However most watermarking systems are frame-oriented and do not focus on the protection of semantic regions. As a result, they fail to protect semantic content, especially in case of the copy-paste attack. In this framework, a novel unsupervised semantic region watermark encoding scheme is proposed. The proposed scheme is applied to real and artwork human objects, localized by two different face and body detection methods. Next, an invariant method is designed, based on Hu moments, for properly encoding the watermark information into each semantic region, using a chaotic pseudo-random number generator. Finally, experiments are carried out, to illustrate the advantages of the proposed scheme, such as:(a) robustness to RST, copy-paste and other attacks, and (b) low overhead transmission.