Computer Vision and Pattern Recognition

Some of our past work includes developing methods for fitting quadric surfaces in range images, for detection military vehicles in laser radar images and for inspecting indutrial castings in range images. We have also developed a parallel technique for 3D object recognition in range images. Many of these methods have exploited models in recognition. We have also developed methods for some 2D and volumetric data sources. Application areas include medical, manufacturing, military and earth and space data.

Aurora Detection
One topic we have investigated involves considering a large archive of satellite imagery of the aurora. In this archive, we aim to identify the location and type of aurorae. One strategy we have considered for this task is using elliptic and related higher order shapes in auroral oval segmentation and object detection. As a result, we have developed a method for auroral oval detection that is superior to the prior methods. In addition, we have designed a new ellipse detection method that is based on least-squares fitting and randomized Hough transform (RHT). An extension of the ellipse detection method to ellipsoid detection has also been studied. The ellipsoid detection method exploits geometric properties of ellipsoids in an evidence accumulation process also based on RHT. We have also investigated steps that can allow deformable ellipsoids to be localized. An on-line image retrieval tool that uses our aurora detection for some of its operations is located at:

Model-Based Surface Classification
We have developed a suite of methods for model-driven identification of quadric surfaces in segmented range imagery. The model-driven parameter estimation is advantageous because it reduces the number and/or range of the parameters to be estimated. The methods we developed seem reasonably immune to the effects of noise and produce good classifications of the images tested. We are currently investigating parallel algorithms to speed up the classification techniques.

3D Object Recognition Using Interpretation Tree Search on a MIMD Machine
Automatic identification and localization of 3D objects in images is a major research issue in computer vision. This problem has been tackled using many approaches, most of which use some form of search technique for matching objects to models. One classic matching paradigm is the interpretation tree (IT) search for matching observed scene features to known model features. Previously, we developed and analyzed parallel implementations of interpretation tree (IT) search for recognition of three-dimensional objects in range images.

CAD-Based Inspection
Currently, many manufactured parts and assemblies are designed using computer-aided design (CAD) tools. Integrated manufacturing enterprises aim to use CAD data in other, downstream functions, such as inspection. Previously, we developed a system that utilizes the CAD model for inspection of iron castings at several points during the assembly sequence. The techniques that were developed were aimed at detecting defects that commonly occur in the casting process. Inspection of castings is a challenging problem because the unmachined castings contain natural roughness from the sand molds into which the metal was poured during casting and because of scratches from collisions with other castings during manufacture. Some of the most common surface defects in castings are pits, excess material, and gross insufficient material. The framework that was designed allowed inspecting for these defects in castings composed of planar and cylindrical surfaces (which includes many basic castings). Automated inspection techniques for several other design features (and their adherence to dimensional tolerances specified in CAD models of castings) were also developed. The casting inspection techniques used template-matching and model-based surface classification methods. The techniques exhibited robust behavior and correctly classified approximately 95% of the castings inspected.

See the technical report.
Recognition of Structures in SAR and LADAR.

Recognition of the cardiac left ventricle
We have developed schemes for object recognition (esp., vehicle recognition) in SAR and LADAR imagery. One of the important measures used in assessing heart function is the ejection fraction (EF). EF is the ratio of the volume of blood ejected from the Left Ventricle (LV) during contraction (systole) to the volume of blood in the LV at the end of dilation (diastole). Other useful measures involve the characterization of the size and shape of the LV and LV wall motion and thickening. Clinically, it is particularly useful to be able to accurately estimate size and shape characteristics throughout the systolic cycle. We have developed a 3 step automatic extraction process for the Left Ventricle on a sequence of gated blood pool (GBP) single photon emission computed topography (SPECT) images. The first step is edge detection, performed using a combined Isotropic and Marr-Hildreth edge detection. The second step is to extract an ellipse-shaped region in each longitudinal slice. The last step is to group ellipse regions. As each ellipse is a potential cross-section of a 3D ellipsoid, the goal of the step is to group related ellipse parameter vectors into classes which represent ellipsoids in the data. In our work, we have used agglomerative hierarchical clustering to collect the ellipse vectors into ellipsoids. The two first steps are automatic (they do not require any human interaction). The last one has highlighted the need for the designer to control, interact, and understand how the single-link clustering properties are mapped on the set of ellipses. To assist this process, we developed a visualization tool that presents a graph that represents the single link clustering. It provides the operator with a scope of commands that allows him to arrange a suitable segmentation from the single-link clustering based algorithm.

Check out the sample images about part of our cardiac functional visualization work!

Parallel Data Mining
We have previously investigated parallel data mining with researchers in the Information Technology and Systems Center. We succeeded in parallelizing the ITSC's ADAM miner using coarse granularity parallelism on networks of workstations and using finer granularity parallelism on an IBM POWER parallel machine.