multi-level image sampling and interpretation: MIST methodology
(Michalski, Duric, Zhang, Maloof)
The goal of this project is to develop a system that can learn descriptions of visual objects (images, visual sources, visual scenes) and to use these descriptions to recognize unknown objects. We have developed a general methodology for this purpose, called multi-level image sampling and interpretation (MIST).
The basic idea under this project can be explained as follows. Given an image with labeled samples of different surfaces, the learning system determines a sequence of operators that transform the image to a “symbolic” image, in which picture elements are labels of corresponding surface areas. The sequence of operators that produces such a labeling serves as a surface description (“surface signature”). A surface description is a logical expression in disjunctive normal form associated with a decision class (here, a texture class). Each conjunction in this expression together with the associated decision class can be viewed as a single decision rule. The basic operator in the process of generating surface description is an application of a set of logic-style rules to transformed surface samples. The rules can be applied in parallel, and serve as “logical templates” that are applied to “events” (attribute vectors) representing surface samples. To recognize an unknown surface sample, the system matches it with all candidate surface descriptions. This is done by applying decision rules to the events in the sample. For each event, the class membership (surface class) is determined. The assignment of the sample to a given decision class (surface) is based on determining which of the candidate classes gets the majority (or) plurality of votes. Thus, even if some events in the sample are incorrectly recognized, the classification of the sample may be correct.
A series of experiments is conducted with gradually increased complexity of data, increased influence of noise, and under variety of other external conditions. The images used for experiments are divided into two groups: office environment images, and outdoor scene images.
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.
Bala, J. and Michalski, R.S., “Learning Texture Concepts Through Multilevel Symbolic Transformations,” Proceedings of the Third International Conference on Tools for Artificial Intelligence, San Jose, CA, November 9-14, 1991.
Channic, T., “TEXPERT: An Application of Machine Learning to Texture Recognition,” M.S. Thesis, University of Illinois, Urbana-Champaign, 1988.
For more references, see publications section.