Research Projects

	Advanced Object Recognition The dominant paradigm for object recognition in computer vision is the model-based approach, wherein features extracted from the image are compared (via a suitable search algorithm, indexing structure, or database query) to comparable features describing the set of objects in the vision system's database (this paradigm is broad enough to include recognition for industrial inspection, recognition of faces, recognition of fingerprints, and recognition of terrain).  This research addresses two of the most formidable barriers to the deployment of model-based object recognition systems in realistic contexts: the complexity of the objects that can be recognized and the scalability of the system to unstructured environments containing hundreds or thousands of different objects.
	Advanced Perceptual Organization The rate at which air- and space-borne platforms produce images far exceeds our ability to process and understand them.  To use these data effectively requires the development of robust computational techniques to extract and classify ground features, and to chart their change with time, subject to wide variations in scene content and imaging conditions.  The complexity of such problems and the need for adaptability suggests the use of machine learning approaches.  Between low-level photometric feature extraction (e.g. edge detection) and object recognition lies a vast, understudied 'gray area' of intermediate-level processing in which the low-level primitives must be grouped and organized into perceptually significant entities.  This research extends and merges prior work in Bayesian and graph-theoretical methods for perceptual organization in complex images with the integration of Bayesian networks and numerical-symbolic learning methods. The challenge is to extract and classify "structures" of features in imagery, as well as develop appropriate organizations for such structures in new image modalities. An example of such a structure is the organization of roads in an aerial image. By exploiting known relationships among the intensity edges in such imagery, we can automate partially or completely the extraction of such structures. This work has applications in a wide range of areas including automated mapping, navigation, and object modeling.
	Intelligent Image and Video Compression The compression of still and moving imagery has received a great deal of attention in the signal processing community, driven by the explosive growth of the Internet, video conferencing, and multimedia applications.  Postprocessing and preprocessing techniques have been developed for the suppression of artifacts in color images encoded at low rates (0.25b/pel and below) by a wavelet transform based image coder, and is currently researching object-based motion estimation for efcient image coding.  Techniques for the intelligent compression of digital echocardiographic image sequences and the view-independent extraction and characterization of surfaces in 3D data have also been investigated.  Wavelet and other transform-based methods hold promise as the basis of 'intelligent' compression algorithms - approaches that use techniques from computer vision first to analyze various aspects of the image structure (including range images and elevation models).  Knowledge of the image structure combined with domain knowledge can be used to advantage in identifying those regions in which the preservation of fidelity is most crucial. The challenge is to develop efficient and optimized compressors for specific types of imagery. Current projects address the intelligent (model-guided) compression of echocardiographic imagery, methods for suppression of compression artifacts in low-rate compressed images, and extension of zerotree-based wavelet coding to video streams. Application areas include multimedia systems and image database processing.
	Dynamic Scene Understanding The challenge is to process and understand images with multiple moving entities. Components of this project include the construction of trackers to monitor single individuals as well as texture analyzers to monitor groups (crowds, flocks, or mobs) of entities on a global (rather than individual) basis. Applications include security threat assessment, driver attentiveness determination, and traffic flow characterization.
	Medical Image Analysis Techniques for improving the quality of images produced by an optical sonography image sensor and reducing the speckle microstructure of sonograms have been developed.  Recent research has also addressed the application of statistical pattern classification techniques (parametric classifiers) to full-wave inverse imaging of the breast; the intended application is the screening of young women for minimal tumors which, if detected early enough, may be removed with minimal invasion and higher probability of complete recovery.  Additional recent research has addressed the detection of boundaries in optical coherence tomography imagery.