Graph Neural Networks (GNNs) are a recently proposed connectionist model that extends previous neural methods to structured domains. GNNs can be applied on datasets that contain very general types of graphs and, under mild hypotheses, they have been proven to be universal approximators on graphical domains. Whereas most of the common approaches to graphs processing are based on a preliminary phase that maps each graph onto a simpler data type, like a vector or a sequence of reals, GNNs have the ability to directly process input graphs, thus embedding their connectivity into the processing scheme. In this paper, the main theoretical properties of GNNs are briefly reviewed and they are proposed as a tool for object localization. An experimentation has been carried out on the task of locating the face of a popular Walt Disney character in comic covers. In the dataset the character is shown in a number of different poses, often in cluttered backgrounds, and in high variety of colors. The proposed learning framework provides a way to deal with complex data arising from image segmentation process, without exploiting any prior knowledge on the dataset. The results are very encouraging, prove the viability of the method and the effectiveness of the structural representation of images.