An image retrieval system providing effective access to desired data in an image database has been studied. Since images are more ambiguous than language, image retrieval that uses only a keyword as a retrieval key is not enough. In this paper, we propose an image retriaval method that uses a rough sketch as the retrieval key. An image in a database system is divided into segments and is described by color, position, size, shape, and geographical relation of each area. These structural features are represented by attributed relational graphs. Similarity is measured by graph matching between stored graphs and a graph derived from a rough sketch.