In this paper, we presented a semantically structured image database for content-based image retrieval. A class descriptor is proposed to represent each class using a multiprototype model, which can be obtained by using a learning scheme, such as the Unsupervised Optimal Fuzzy Clustering algorithm, on a group of sample images manually selected from the class. Based on the proposed Image-Class Matching Distance, a similarity measure at the semantic level between an image and classes, images can be annotated by tokens of classes. Hence, composite features of images, including low-level descriptors, class descriptors, and image annotation, are stored into a structured feature database corresponding to the semantically structured image database. From experiments, it can be concluded that the performance of the semantically structured CBIR system is improved greatly in terms of retrieval time and efficiency.