The project arises the solution of the Semantic Image Segmentation through the establishment of a machine learning method to predict complex structures within the image such as graphs and trees in which are defined relationships or distances between semantic objects present on it. These entities are bounded within a range of possible configurations that can produce scenes and emerging concepts. The implemented method should therefore keep a set of labeled training images that determining processing base objects through a basic segmentation, so the possible output set is reduced by defining a range of looseness for the discovery of new relationships and concepts. As a result of structure image prediction can be established a set of automatic annotation processes, recovery of similar entities or associated with a particular concept and finally abstraction of relationships and patterns that define new and different concepts in the image.