In this paper we introduce a new abstract image representation based on Edge Pixel Neighborhood Information (EPNI). It is applied in image retrieval problem when user query is a fast drawn, rough example. The representation consists of two main elements. A neighborhood vector f and a vicinity table v. The former contains the frequencies of edge pixels with similar directions and the latter holds information about neighboring edge directions. An image similarity measure based on EPNI components is also designed and compared with some other measures known from the literature. Experimental results show a good recognition accuracy in a data set containing a wide range of color images.