A Theoretical Comparison of Statistical Feature Selection Criteria for Realtime Pattern Recognition.
Abstract
Tight error bounds on the information measure for feature selection are obtained and shown to be better than those offered by the Bhattacharyya distance. Feature subset constructed from features ordered by the information measure is also shown to perform better. The analytical study is supported by computer simulation on handprinted characters. To meet the realtime recognition needs, the computational complexity in feature selection is considered. For features with Markov dependence, recursive computation of mutual information is available such that the computational complexity increases only linearly while the performance could be much better than available with the independent assumption. For n Markov dependent binary variables, a saving of (2 sup n)/4n in the computation time for feature subset construction is available. Other distance measures are also considered. It is concluded that the information measure is the best compromise choice as far as both the error performance and the computational complexity are concerned. Application of the theoretical study to real seismic data was started near the end of the grant period. Some preliminary analysis of the dat data is described in this report. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 20, 1973
- Accession Number
- AD0757532
Entities
People
- Chi-hau Chen
Organizations
- University of Massachusetts Dartmouth