Multiclass AdaBoost Classifier Parameter Adaptation for Pattern Recognition - Publikacja - MOST Wiedzy

Wyszukiwarka

Multiclass AdaBoost Classifier Parameter Adaptation for Pattern Recognition

Abstrakt

The article presents the problem of parameter value selection of the multiclass ``one against all'' approach of an AdaBoost algorithm in tasks of object recognition based on two-dimensional graphical images. AdaBoost classifier with Haar features is still used in mobile devices due to the processing speed in contrast to other methods like deep learning or SVM but its main drawback is the need to assembly the results of binary two-class classifiers in recognition problems. In this paper an original method of selecting the parameter values of the assembling algorithm using many similar face recognition tasks is proposed. The parameter optimization is done by checking all possible vectors of parameter values. The recognition results with optimized parameter values is $10\%$ better in 8-class face database famous48\footnote{http://eti.pg.edu.pl/documents/176468/27493127/famous48.zip} tasks than using random heuristic which can be represented by the average of all possible vectors of parameter values.

Cytowania

0
CrossRef
0
Web of Science
0
Scopus

Jerzy Dembski. (2017). Multiclass AdaBoost Classifier Parameter Adaptation for Pattern Recognition, 525(Chapter 24), 203-210. https://doi.org/10.1007/978-3-319-47274-4_24

Informacje szczegółowe

Kategoria:
Inna publikacyjna praca zbiorowa (w tym materiały konferencyjne)
Typ:
publikacja w wydawnictwie zbiorowym recenzowanym (także w materiałach konferencyjnych)
Opublikowano w:
Advances in Intelligent Systems and Computing nr 525, strony 203 - 210,
ISSN: 2194-5357
Tytuł wydania:
Image Processing and Communications Challenges 8 strony 203 - 210
Język:
angielski
Rok wydania:
2017
Opis bibliograficzny:
Dembski J.: Multiclass AdaBoost Classifier Parameter Adaptation for Pattern Recognition// Image Processing and Communications Challenges 8/ ed. Ryszard s. Choraś : Springer, 2017, s.203-210

wyświetlono 18 razy

Meta Tagi