読唇による母音の自動認識

概要

論文の詳細を見る
This paper describes a method of recognition of five vowels by machine lip reading. The feature values of height and width of lip shape and bottom position of jaw were extracted. Compared to the feature values of only lip shape, adding jaw information revealed a significant difference of discrimination. By introducing a rule based on the Mahalanobis' generalized square distance and the maximum likelihood discrimination function, rejecting the recognition of ambiguous feature values of a vowel and instructing the subject to reenter the vowel decreased the failed recognition rate. Finally, an instruction system which recognies by successively adding the reentered feature values to statistics of the maximum likelihood discrimination function was developed. Using this system, the vowels were correctly discriminated by 100 %.
社団法人日本機械学会の論文
1987-12-25