|
The application of pattern recognition techniques to the data from the Faint Images of the Radio Sky at Twenty Centimeters Survey (FIRST) is explored. The methodology developed for the automatic classification of a particular type of radio-source morphology, the so-called "bent doubles" is described. While the procedures and techniques for doing supervised pattern recognition using accurately classified training sets, high-resolution images, and determinate shapes are well established, there exists little guidance for dealing with more ambiguous populations. The implications of low resolution for pattern recognition are discussed and methods for dealing with more ambiguous populations are presented. Bent doubles with three components are considered. With approximately 6% of the catalog sources being a component of a triple, visual classification suggests that approximately 6% of triples are bent. In addition to the problems associated with low resolution, the classification task is complicated by the lack of scale and orientation information, the chance superposition of sources, the very ambiguous nature of the visual classifications, and the considerable variation in bent morphology. For a bent or ambiguous or nonbent visual classification, classifiers are presented that produce results having up to an 80% recognition rate when looking at the top-ranked 16% of sources. The associated nonambiguous false positive rate varies between 14 and 18%. Given that recognition rates cannot be expected to be better than the accuracy of the training set, this is believed to be within a few percent of the maximum recognition rates attainable without an observationally verified training set. Decision tree and neural network results are compared and show, within associated errors, substantially equivalent results. Ranking of sources is demonstrated and the highest-ranked sources are in good agreement with visual classifications. In addition to providing determinate expressions for characterizing the target class, the resulting distributions suggest that the automated classifications constrain the target class in a better manner than the visual classifications.
|
