An object-oriented approach for semantic-based image retrieval is presented. The goal is to identify key patterns of specific objects in the training data and to use them as object signature. Two important aspects of semantic-based image retrieval are considered: retrieval of images containing a given semantic concept and fusion of different low-level features. The proposed approach splits the image into elementary image blocks to obtain block regions close in shape to the objects of interest. A multiobjective optimization technique is used to find a suitable multidescriptor space in which several low-level image primitives can be fused. The visual primitives are combined according to a concept-specific metric, which is learned from representative blocks or training data. The optimal linear combination of single descriptor metrics is estimated by applying the Pareto archived evolution strategy. An empirical assessment of the proposed technique was conducted to validate its performance with natural images.