Discrimination of images is necessary in many tasks, either understanding or indexing for example. Here we are concerned by indexing. More precisely we are working about initial letters extracted from early Renaissance printed documents as an application. Different observation levels can be considered according to the applications, either details can be observed or more globally what could be called the style. Here we are concerned with a global view of image. Then we are going to present a new method to index ornamental letters in ancient books. We show how the Zipf law, originally used in mono-dimensional domains can be adapted to the image domain. We use it as a model to characterize the distribution of patterns occurring in these special images that are initial letters. Based on this model some new features are extracted and we show their efficiency for image indexing and retrieval.