machine vision and learning

(Michalski, Duric, Zhang, Maloof, Bloedorn)

The goal of this project is to develop methods and experimental vision systems that are capable of learning general visual concept descriptions from specific observed objects, and then use these descriptions to efficiently recognize new objects in a visual scene. It is assumed that the system should be able to recognize objects among other objects in a scene under a variety of conditions, such as changing viewpoints, changing illumination, object overlap, and in the presence of noise in the sensory data.

Our approach is based on a two-pronged architecture in which the first prong processes the surface information about objects, and the second prong processes the shape information. Learning unique surface characteristics (“surface signatures”) involves problem-oriented transformations of the representation space, and an iterative application of an inductive learning program. The input to the system are classified samples of surfaces.

In the object recognition phase, the system applies the learned rules to identify the surface, and then uses this information to generate a set of candidate hypotheses about the object’s identity.

These hypotheses are then employed to retrieve specific 3D structural models of the objects from a knowledge base. We use CAD/CAM descriptions of objects to discover their characteristic structural and symbolic features and feature relations, and to learn recognition strategies. These processes are also driven by vision tasks, such as localization, recognition and inspection. Learned models are then used to determine characteristics of objects sufficient to identify the object in the scene. These discriminatory characteristics are determined by a process called dynamic recognition.

The research on this project is conducted in collaboration with the Computer Vision Laboratory of the University of Maryland.

Our laboratory, in collaboration with the UMD Computer Vision Laboratory, organized the NSF/ARPA workshop on Machine Vision and Learning . The workshop was held in Harpers Ferry, WV, in October 15-17, 1992. It was the first workshop that brought together leading researchers in computer vision and machine learning. Below is a reference to the report that was partially based on this workshop and partially based on new material:

Selected References

Michalski R.S., Rosenfeld, A., Aloimonos Y., Duric, Z., Maloof M.A., Zhang Q., “Progress On Vision Through Learning: A Collaborative Effort of George Mason University and University of Maryland, Proceedings of the Image Understanding Workshop, Palm Springs, CA, Feburary, 1996.

Michalski, R.S., Zhang, Q., Maloof, M.A. and Bloedorn. E., The MIST Methodology and its Application to Natural Scene Interpretation, Proceedings of the Image Understanding Workshop, Palm Springs, CA, Feburary, 1996.

Zhang, Q., Duric Z.,Maloof, M.A. and Michalski, R.S., “Target detection in SAR images using the MIST/AQ method” Reports of Machine Learning and Inference Laboratory, MLI 96-12, George Mason University, Fairfax, VA, 1996.

Michalski, R.S., Rosenfeld, A., and Aloimonos, Y., “Machine Vision and Learning: Research Issues and Directions,” Reports of the Machine Learning and Inference Laboratory, MLI 94-6, Machine Learning and Inference Laboratory, George Mason University, Fairfax, VA; Reports of the Center for Automation Research CAR-TR-739, CS-TR-3358, University of Maryland, College Park, MD, October 1994.

For more references, see publications section.