"Semantic gap" is an open challenging problem in content-based image retrieval. It rejects the discrepancy between low-level imagery features used by the retrieval algorithm and high-level concepts required by system users. This paper introduces a novel image retrieval scheme, CLUster-based rEtrieval of images by unsupervised learning (CLUE), to tackle the semantic gap problem. CLUE is built on a hypothesis that images of the same semantics tend to be clustered. It attempts to narrow the semantic gap by retrieving image clusters based on not only the feature similarity of images to the query, but also how images are similar to each other. CLUE has been tested using examples from a database of about 60,000 general-purpose images. Empirical results demonstrate the effectiveness of CLUE.