In the last years unsolicited commercial emails (UCE), commonly known as spam emails are becoming an increasing problem for the Internet communities. Many strategies have been proposed, although we are still far away from a satisfactory and fundamental solution. In this paper we describe a novel method for detecting spam messages analyzing both text and image attached components. In particular, we describe an architecture that can overcome some problems that are still boarded on the state-of-the-art spam-filters. This approach takes into account some techniques that are able to get the semantic richness of natural language and some features given from the recent spam evolution based on images. Eventually, we describe our experimental settings and results together with a comparison of our method with existing tools.