An image can be considered as a combination of regions, which have some intrinsic homogeneities in color or texture, etc. The shapes in an image are described by the boundaries of these regions and their spatial relations. To retrieve image by shape, it is a key problem to measure the similarity among shapes and match their spatial relations. Adopting template-matching technique, the authors present a computation method for simple shape similarity. Based on the two-dimensional set string representation of the spatial relation of these simple shapes, a spatial relation match algorithm is also presented. The whole retrieval process is divided in three stages, i.e., the rough stage, the refining stage and the spatial relation matching stage. The experiment shows this method is rapid and accurate.