Системи та методики обробки інформації

Permanent URI for this collection

https://elar.naiau.kiev.ua/handle/123456789/25536

Browse

Browsing Системи та методики обробки інформації by Author "Kunakh, N.I."

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Оцінка ефективності штучних нейронних мереж. EVALUATION OF THE EFFECTIVENESS OF ARTIFICIAL NEURAL NETWORKS
    (2018) Кунах, Н.І.; Kunakh, N.I.; Ткаленко, О.М.; Tkalenko, O.M.; Харлай, Л.О.; Kharlai, L.O.
    The purpose of the article is the use of artificial neural networks to improve the quality of images when increasing their permission. To achieve the goal, various types of most promising artificial neural networks were considered – mathematical models that describe the system of artificial neurons interconnected among them, as well as the implementation of these models. The analysis of modern interpolation algorithms, which include the initialization of paired points by the values of the reduced image, the metrics for assessing the quality of the algorithms; using artificial neural means to increase image resolution; methods of increasing the efficiency of artificial neural networks. A software implementation has been created using several interpolation algorithms and models of artificial neural networks. An analysis of modern approaches to solving the problem of increasing the resolution of images has shown that despite the existence of relatively fast and high-quality algorithms, new methods are constantly emerging. In recent years, artificial neural networks are increasingly used to solve this problem. In most studies, different implementations of the models of the curtain neural networks are used. The algorithms of bicubic interpolation, Lancosha filter, models of artificial neural networks SRCNN and SRGAN were studied in this work. The choice of interpolation algorithms is influenced by the fact that they are quite effective in terms of the ratio of the time of work and the quality of the result. Also, the choice is influenced by the fact that these algorithms are widespread and used in many applications. When comparing results with objective metrics, it can be noted that generative- competitive model copes with image processing more efficiently than the creeping neural network. But adapting such a model requires much more time. Using the generic-competitive network model to handle test images also takes a little more time than the rollout. But the result, which allows both models, far outperforms interpolation algorithms.