(20) * << * >> * Russian * English * Content * All Issues
Pages: 140-142.
Full text of article: Russian language.
Abstract:
tantiation of the feature space dimension is one of the main problems in the general problem of pattern recognition learning. The work [1] uses the measures based on the calculation of entropy to teach the digit recognition. The use of these measures requires a large number (ensemble) of realizations to identify the processing patterns. In the work [2], to select the feature space dimension with a small number of training objects, it was proposed to use the measures of the feature vectors orientation relative to the null space of the transposed feature matrix. These measures essentially characterize the multicollinearity of the column vectors, which form the feature matrix. In this paper, the analysis procedure proposed in [2] is applied to the problem of formal analysis of the feature space in the problem considered in [1]. Therein, the area of localization of digits was also changed as compared with [1]. In particular, when choosing this area, it was taken into account that the vertical size of digits is usually larger than the horizontal size.
Citation:
Fursov VA, Shustov VA. Formation of feature space by the criterion of conjugacy of measurement vectors. Computer Optics 2000; 20: 140-142.
References:
© 2009, IPSI RAS
151, Molodogvardeiskaya str., Samara, 443001, Russia; E-mail: journal@computeroptics.ru ; Tel: +7 (846) 242-41-24 (Executive secretary), +7 (846) 332-56-22 (Issuing editor), Fax: +7 (846) 332-56-20