This PhD. thesis addresses the unsupervised hierarchical segmentation of images and the automatic extraction of the image partitions providing the most semantically relevant explanations of the image at different levels of analysis. From a semantic and practical perspective, image segmentation is a first and key step for image analysis and pattern recognition since region-based image representations provide a first level of abstraction and a reduction of the number of primitives, leading to a more robust estimation of parameters and descriptors.
The proposal of this dissertation is based on an important class of hierarchical bottomup segmentation approaches, known as region merging techniques. These approaches naturally provide a bottom-up hierarchy, more suitable when no a priori information about the image is available, and an excellent compromise between efficiency of computation and representation.
The first part of the dissertation is devoted to the proposal, under a statistical framework, of a family of unsupervised region merging techniques. These techniques are characterized by general and non-parametric region models, with neither color nor texture homogeneity assumptions, and a set of innovative merging criteria, based on information theory statistical measures. The scale consistency of the partitions is assured through (i) a size regularization term into the merging criteria and a classical merging order, or (ii) using a novel scale-based merging order to avoid the region size homogeneity imposed by the use of a size regularization term. Moreover, a partition significance index is defined to automatically determine the subset of most representative partitions from the created hierarchy. Most significant automatically extracted partitions show the ability to represent the semantic content of the image. Results are promising, outperforming in most indicators both object-oriented and texture state-of-the-art segmentation techniques.
The second part of the thesis is focused on the fusion of hierarchical segmentation results, obtained by different segmentation techniques or from different information channels of the same image, with the purpose of increasing the robustness and accuracy of the final solution. In this case, a generic and scalable segmentation scheme based on a cooperative principle, named cooperative region merging, is designed to combine in an unsupervised manner a set of hierarchical region-based representations. The intuition behind this cooperative approach is to iteratively establish a basic or conservative consensus between the independent techniques that can be used as the starting point from which further consensus may be built, similar to a negotiation process for decision making. In addition to the new fused hierarchy of partitions, the proposed scheme automatically provides a subset of partitions considered most relevant from the fusion viewpoint. The combination of hierarchical segmentation results applying different segmentation techniques over the same color image leads to a global improvement of the accuracy and the stability of the segmentation results. Moreover, the use of the cooperative approach for the fusion of segmentation results from heterogenous information
channels is presented. Application examples demonstrate the high flexibility and potentiality of the cooperative region merging scheme into a wide range of applications and fusion problems (for instance, in multiview processing and remote sensing). The proposed fusion strategies are able to naturally incorporate a priori available knowledge on the types of information to combine, or on the specificities of the particular fusion problem, and to improve the efficiency and reduce the computational load of the fusion process without compromising the accuracy of the segmentation results.