Full text of article: Russian language.
This paper presents a problem of reducing the dimensionality of a feature space in recognition problems on images and proposes a certain problem-solving technique. The proposed method allows us to reduce a number of features required to solve a specific recognition problem, from several hundred thousand of features (number of pixels of the original image) to a few tens or hundreds of features. The developed method consists of three stages. At the first stage, we calculate two-dimensional maps of features from a training sample for each image (for example, spatial filters processing results, spectral features), and in these maps we preselect the features by the total-to-average intraclass variance criterion. Then the selection is performed by searching different combinations (by the method of sequential addition and deletion of features) using the criterion of a specific recognition problem for which the features are selected. At the last stage, selected combinations of the features are tested on a control image sample, and a final decision on selection of a feature set is made in order to use it in recognition. The proposed method was successfully applied in selection of the features for implementation of the test person recognition system by face images on documents.
feture space reduction, intraclass variance criterion, person recognition system, face images.
Glumov N.I., Myasnikov E.V. Method of the informative features selection on the digital images [In Russian]. Computer Optics 2007; 31(3): 73-76.
© 2009, ИСОИ РАН
Россия, 443001, Самара, ул. Молодогвардейская, 151; электронная почта: firstname.lastname@example.org ; тел: +7 (846) 332-56-22, факс: +7 (846 2) 332-56-20