Probabilistic multifeature/multisensor integration for automatic object recognition

This dissertation describes a new methodology for detecting and recognizing manmade objects in visual images, second generation Forward Looking Infrared (FLIR) images, and registered multisensor images for Automatic Object Recognition.; 3D and 2D object recognition from images has been a subject of study in the computer vision community since the mid-70's. A plethora of systems and methodologies has been proposed in an attempt to find a solution to the problem. However, available solutions are not versatile in that they work only in constrained or situation-specific environments. A system that uses information from multiple features or sensors can employ redundancy, diversity and complementarity to overcome the shortcomings of single-sensor systems and improve performance. In this dissertation, a general multifeature/multisensor framework is proposed which does not simply expand the dimensionality of the feature space, but which can discern new features to provide greater discrimination. Using this framework, a more focused methodology is described for localization of manmade objects in complex scenes by learning multiple feature models in images. The methodology is based on a modular structure consisting of multiple classifiers, each of which solves the problem independently based on its input observations. A higher level decision integration is obtained through a supra-Bayesian scheme. Finally, a recognition methodology is proposed to classify segmented object regions by extending the multifeature/multisensor framework. Recognition is performed hierarchically by using geometric and photometric features for object representation.; This dissertation also presents a methodology for evaluating the performance of classification systems and interpreting the usefulness of various features. The methodology is based on using neural and Bayesian classifiers, and extracting rules from them to form a new rule-based system. This is a novel and useful method for reasoning about the performance of classifier systems.; Finally, a method based on texture analysis for detection and segmentation is also presented which uses Bayesian formulation for labeling similar regions. Similarity is defined via texture features obtained by Gabor wavelets. Multivariate Gaussian distributions are employed to model the feature class-conditional densities, while the Markov process is used to characterize the distributions of the region labeling due to each feature.