|
The aim of the article is to analyze and compare the performance and accuracy of
architectures with a different number of parameters on the example of a set of
handwritten Latin characters from the Polish Handwritten Characters Database
(PHCD). It is a database of handwriting scans containing letters of the Latin alphabet
as well as diacritics characteristic of the Polish language. Each class in the PHCD
dataset contains 6,000 scans for each character. The research was carried out on six
proposed architectures and compared with the architecture from the literature. Each
of the models was trained for 50 epochs, and then the accuracy of prediction was
measured on a separate test set. The experiment thus constructed was repeated 20 times
for each model. Accuracy, number of parameters and number of floating-point
operations performed by the network were compared. The research was conducted on
subsets such as uppercase letters, lowercase letters, lowercase letters with diacritics,
and a subset of all available characters. The relationship between the number of
parameters and the accuracy of the model was indicated. Among the examined
architectures, those that significantly improved the prediction accuracy at the expense
of a larger network size were selected, and a network with a similar prediction
accuracy as the base one, but with twice as many model parameters was selected.
|
