The classical segmentation of images is basically subdivided into three main techniques: global, which is based on the relative knowledge about the intensity of the components from the pixels; region-based, which consists of dividing the images in similar regions; and edge-based, which consists of edge detection to build closed lines around the detected object in the image. This work describes a different approach to apply the segmentation of images: the Markovian Geometric Diffusion (MGD). The MDG enables the extraction of the representativeness of the vertices from a graph, which propitiates the elimination of the elements from the graph that hold few representativeness. In the context of digital images, the image is converted into a graph, and the MGD is intended for pixels, being representativeness-based from each of them. The method developed in this work builds descriptors for objects by eliminating few relevant information from the image, generating as result a segmented image