Open AccessSelection of appropriate architecture and parameters of neural network for images recognition and classification
Author(s) -
В. В. Кукарцев,
А. С. Михалев,
A V Tarasevich,
V. S. Tynchenko,
A.R. Ogol,
V V Khramkov
Publication year - 2019
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1399/3/033105
Subject(s) - convolutional neural network , computer science , artificial intelligence , task (project management) , artificial neural network , pattern recognition (psychology) , image (mathematics) , selection (genetic algorithm) , contextual image classification , sample (material) , table (database) , machine learning , data mining , engineering , chemistry , systems engineering , chromatography
In this article the image recognition and classification problem is considered. Further, there will be suggestion of the new way of using a convolutional neural network to solve the considering problem in an example of a particular task, such as the recognition of ischemic stroke on magnetic resonance images. Moreover, it is considering reasons of choosing the image classification and making arguments obtained by a literary analysis of studies affecting the task. Authors have investigated different neural networks architectures and the accuracy depending on training speed, numbers of layers, numbers of epochs and mini sample size. Experiments result is presented in the table form. During the studying, magnetic resonance images of different diseases having same signals with a considering pathology in diffusion-weighted images format has been selected for the training of the convolutional neural network and for getting the most favourable result. In the result, a suitable architecture with defined parameters has been selected to get the highest accuracy.