Relevance feedback which is used in content-based image retrieval (CBIR) has been considered as the efficient technique to improve the retrieval performances. The traditional relevance feedback technique demonstrates a disability to use the users' historical feedback information sufficiently gotten together by the system in the former retrieval processes when initiating a new query session. In this paper, an approach to relevance feedback based on long-term learning strategy using the historical retrieval information is presented for the content-based image similarity retrieval. The approach adopts a semantic covering set constructed dynamically to deposit the users' historical retrieval information produced in previous retrieval processes, and predicts the semantic correlation between the images in database and query sample according to the historical retrieval information when carrying out a new query session. The performance of an experimental image retrieval system using this approach is evaluated on a database of around 3000 images. Empirical results demonstrate improved performances compared with the CBIR system with the traditional relevance feedback technique using the same image similarity measure.