Método interactivo multiescala de segmentación de imágenes

In the present work, an interactive method for segmenting images is proposed. This method is based on a multi-scale approach that permits going from finer scales to coarser ones (also from coarser scales to finer ones). The main goal of the proposal consists in proposing a simple method that allows carrying out a segmentation of the original image in a simple way. Thus, user with a minimum of knowledge in image processing can compute a segmentation of the image. The method is based on a minimum spanning tree structure (MST) which is built during immersion process of the image to compute a catchment basins. Each catchment basin is associated with a given node in the MST, where the edges of the MST graph are weighted according to some measures (high, volume, area) that form the criteria to extract the different segmentations of the image. A pre-processing step is carried out on the image in order to simplify image segmentation process. After the MST is completed, a set of transformations can be applied on the MST structure. Working with the MST structure has an important advantage because of the small number of nodes to be processed in the MST structure with regard to the information of the image. The two principal transformations working on the graph are the split and merge operations. The changes on the MST by the splitting and merging nodes will imply the same changes on the image by merging and splitting the associated regions. The first operation called split allows us to compute a set of sub-graphs according to the regions required by the user. Each region of the segmented image will be associated to a sub-graph. In a recursive way the user can split again each sub-graph in another set of sub-graph. On the other hand, the second operation permits the merging two of sub-graphs. In this case, the regions of image to be merged are selected by the user. This operation is only made if there is an edge connecting the sub-graphs on the MST. Both basic operations permit building a set of more complex operations on the graph. The method proposed in the present work was validated with a set of images with different structural characteristics. This validation shows the performance of the proposal.