Content-based image retrieval (CBIR) and semantic-based image retrieval (SBIR) have attracted great research attentions. However, they all have disadvantages. This paper proposes a retrieval method trying to overcome them. To achieve this we first introduce an approach of rough set-based low-level features selection. We propose an approach of feedback-based semantic-level features annotation. We also introduce a corresponding computing technology of similarity. We then build a model of combination of these two kinds of methods. Experimental results show that our method is more user-adaptive, and can achieve better performance compared with another retrieval method which is only based on low-level features.