Improvement of neural network model topology for object segmentation in digital images based on convolutional neural networks

Nowadays, convolutional neural networks have demonstrated significant performance gains over traditional machine learning methods for various real-world computational intelligence tasks such as digital image classification. However, to achieve the best accuracy, the network topology should be modeled using different architectures with different number of filters, kernel size, number of layers, etc., which actualizes the problem of developing and justifying appropriate selection methods. Taking into account the above mentioned, the aim of the paper is to justify an approach that will improve the topology of the neural network model for object segmentation in digital images based on convolutional neural networks. The research methods are system analysis, modeling, machine learning and fuzzy logic theory, and decision-making theory. As a result of the analysis, the paper proposes an algorithm to improve the topology of the neural network model based on differential evolution to optimize the accuracy of image segmentation and the training time of the network. Differential evolution is applied to determine the optimal number of layers in the network topology, which promotes faster convergence. Within the proposed algorithm, an encoding step was identified to represent the structure of each network using a fixed-length integer array, after which it is proposed to utilize differential evolution processes (mutation, recombination, and selection) to efficiently explore the search space. Prospects for further research are to develop methods and techniques to encode a candidate solution using different numbers of hidden blocks in each convolution.