Wyniki wyszukiwania - Biblioteka Nauki

Ten serwis zostanie wyłączony 2025-02-11.

Nowa wersja platformy, zawierająca wyłącznie zasoby pełnotekstowe, jest już dostępna.
Przejdź na https://bibliotekanauki.pl

Ograniczanie wyników

1 Safety and Reliability of Systems and Processes

1 2023

1 Mazurkiewicz J.

Znaleziono wyników: 1

Liczba wyników na stronie

Wyniki wyszukiwania

Wyszukiwano:
w słowach kluczowych: art style recognition

Sortuj według:

Ogranicz wyniki do:

Softcomputing system for painting art style verification

100%

Mazurkiewicz J.

Safety and Reliability of Systems and Processes

2023

tom Summer Safety and Reliability Seminar 2023

157--174

The chapter discusses the foundations for the system to verify and recognize the art style. Such a system seems to be interesting for the first step of the painting fraud identification and following the possible path of the different style influence for the final shape of the masterpiece. The approach presents image recognition using convolutional neural networks. These networks, due to their structure resembling the sight apparatus, and due to their efficiency in the case of two-dimensional data, are very often used for image recognition. Style recognition in art is currently a hot topic in machine learning circles. Courtesy of museums and galleries, there are now many databases available on the web that can be used in scientific work. The classes on which the network was to be tested are styles in history that are associated by the layman with the subject of art. This group includes Renaissance, Baroque, Romanticism, Neoclassicism, Surrealism, Cubism, Art Nouveau, Abstract Expressionism, Pop Art, and Impressionism. These classes were described in the work in terms of parameters that could affect the learning of the neural network. The networks were tested to determine the best parameters for identifying artistic styles. Networks with changing filter values, stride, and pooling parameters, and by selecting various additional layers were tested. The most important parameter was overfitting, which had to be prevented. As a result, networks peaked at 40% in Top1 and 80% in Top3. For smaller data, this result was optimistic for further research in recognizing other parameters, as well as using networks that were previously taught specific characteristics of styles, such as frequently used motifs or colors.