In applications based on template matching, the design of the template is a critical point and has a considerable effect on the overall performance that can be obtained. In this paper we present a method, applicable to binary images, for extracting robust templates. Starting from a provisional prototype (defined as the image that minimizes its overall distance from the samples of a training set), the template is obtained by eliminating unreliable pixels determined by means of an entropy-based criterion. The method is compared experimentally with the matched filtering technique in the recognition of symbols on topographic maps and shows promising results regarding the recognition rate and the computational cost of the matching process.