A Theoretical Comparison of Statistical Feature Selection Criteria for Realtime Pattern Recognition.
Abstract
The paper summarizes the theoretical studies accomplished over a one year period on comparisons of statistical feature selection criteria for realtime pattern recognition. Improved upper and lower error bounds are discussed where entropy criterion are shown superior to B-distance criteria. The relationship between a feature set and its subset is discussed, and the error bounds of a recognition system with a simple rejection feedback strategy are examined. It is also shown that for realtime pattern recognition, it would be best to evaluate the tight upper and lower bounds instead of the error probability itself. The closed form solution for the information criterion still remains to be studied for the B-distance criterion, however, if the Parzen-Murthy nonparametric estimation of the probability density if used, then an upper bound has been derived for the B-distance criterion. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1972
- Accession Number
- AD0746737
Entities
People
- Chia‐Hung Chen
Organizations
- University of Massachusetts Dartmouth