Feature Weighted Minimum Distance Classifier with Multi-class Confidence Estimation

摘要

In many recognition tasks, a simple discrete class label is not sufficient and ranking of the classes is desirable; in others, a numeric score that represents the confidence of class membership for multiple classes is also required. Differential diagnosis in medical domains and terrain classification in surveying are prime examples. The Minimum Distance Classifier is a well-known, simple and efficient scheme for producing multi-class probabilities. However, when features contribute unequally to the classification, noisy and irrelevant features can distort the distance function. We enhance the minimum distance classifier with feature weights leading to the Feature Weighted Minimum Distance classifier. We empirically compare minimum distance classifier and its enhanced feature weighted version with a number of standard classifiers. We also present preliminary results on medical images with acceptable performance and better interpretability.