Electronic Archive Khmelnitskiy National University ELARKHNU
Переглянути архів ІнформаціяMetadata
| Поле | Співвідношення |
| Title | Equalization of nonuniform distribution of recognition errors percentage over classes in classifying shifted monochrome 60-by-80-images |
| Names |
Romanuke, V.V.
Романюк, В.В. |
| Date Issued | 2015 (iso8601) |
| Abstract | A study for equalizing nonuniform distribution of recognition errors percentage over classes is presented. The problem is exemplified on classifying shifted monochrome 60by80images of 26 English alphabet letters. A general conception of adjusting the training process of twolayer perceptron classifier for the equalization is stated. In classifying shifted monochrome 60by80images, this conception is used partially for cutting off extremely long training sample set. Within the example, the nonuniformity is reduced nearly for 25 %. The conception root is that the class representative recognized poorer is repeated in the training set. As the class recognition errors percentage increases, the repeat number is greater. However, the equalization conception is usable only for reasonable number of classes, so that the training sample set length could be shorter. Besides, the twolayer perceptron classifier can be adjusted by four parameters, determining the training sample set length and number of times when this set is passed through the perceptron. While equalizing, the hidden layer size of the twolayer perceptron should not be changed. For estimating nonuniformity, variance unbiased estimates may be used as well. The variance is decreased further using boosting ensembles of perceptron classifiers. |
| Genre | Стаття |
| Topic | object classification |
| Identifier | Romanuke, V. V. Equalization of nonuniform distribution of recognition errors percentage over classes in classifying shifted monochrome 60-by-80-images [Текст] / V. V. Romanuke // Вісник Хмельницького національного університету. Технічні науки. – 2015. – № 2. – С. 146-152. |