Salient structural elements are ubiquitous in natural textures, and their distribution exhibits some stochastic distribution features. Current texture synthesis algorithms can neither preserve the integrity of the elements nor capture this distributive information. We present an algorithm to treat this high-level visual information. Here, we address the issue by taking specific care of the structural elements. Our texture synthesis process grows the target texture one structural element at a time. A Markov random field model is assumed for the distribution of the salient structural elements. In the analysis process, the salient structural elements are first extracted, and their topological relationship is constructed. Next, in the synthesis process, to identify an unknown element with neighbors synthesized, we query the sample texture and find the most similar neighbor. The randomness of the target texture can be improved by randomly transforming the structural elements. Experimentation shows that our results pass scrutiny by the high-level visual processing of humans.