In object recognition systems based on template matching by calculating the distance metric between preassigned templates and given object, the template selection can severely affect the recognition quality. This article presents a novel method for image selection in object recognition systems, which can improve the identification quality. Its uniqueness lies in the fact that the images, what is the best for further recognition, were selected automatically in a way that provides the best quality. The technique is based on a SOM (Self-Organizing Map) and k-mean clustering. In this article, we describe the selection procedure and show the result of its usage in the system.