Techniques for Efficient Feature Representation in Pattern Recognition,
Abstract
Techniques are developed for efficient encoding of features in pattern recognition. These techniques can be used in conjunction with conventional feature selection procedures for optimal representation of patterns. It is shown that the number of binary digits needed to represent patterns can be greatly reduced using efficient feature encoding. Three isolated-word data sets are used to evaluate the techniques in a speech pattern recognition system implemented on a digital computer. The largest data set consists of four thousand utterances based on a forty word vocabulary. A rate-distortion bound and a mutual information measure are used to determine the number of binary digits necessary to represent each feature. On the largest data set the encoding procedures reduce the number of binary digits necessary from approximately four hundred to fifty (a factor of eight) without loss in classification accuracy. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 22, 1970
- Accession Number
- AD0716854
Entities
People
- Joseph Evan Shoenfelt
Organizations
- North Carolina State University