Development of convolutional neural networks for solving classification and localization problems

Kochnev Alexander Alexandrovich

Over the past few years, the use of approaches to the recognition of images, objects, optoelectronic devices designed for visual control or automatic image analysis in combination with computer vision methods based on the use of neural networks have proven their effectiveness in solving various problems. Convolutional networks, being a key element of most data mining systems, are able to influence the processes occurring in various systems due to the relationship between data about a specific event and the ability to predict future events. However, the development in the field of adaptation and use of neural networks for solving localization and classification problems is very slow. And if some of the problems in this area related to object recognition have already been overcome by training neural networks based on elementary algorithms, then such problems as the performance of neural networks and the choice of optimal network training algorithms for these purposes are still not solved. In the article, the author examines the main theoretical problems of the development of convolutional neural networks for solving classification and localization problems and comes to the conclusion that it is necessary to develop approaches to faster and more accurate training of networks. At the same time, such approaches should be comprehensive and focused not only on solving the problem of increasing the amount of processed information without losing the quality of the network, but also on increasing the number of layers of the neural network without losing the accuracy of the neural network and its performance.

Keywords:convolutional neural networks; classification and localization tasks, neural network training, recognition tasks, computer vision