Класифікація соняшника за допомогою згортково-капсульної моделі нейроної мережі CNN-CapsNet

Abstract

Sunflower classification using convolutional- capsular model of neural network CNN-CapsNet of improved architec-ture, which combines the convolutional neural network CNN and the capsular neural network CapsNet and enables to use the advantages of these two architectures is suggested in the paper. The conducted review of the literature sourcesallows to conclude that the advantage of convolutional neural network CNN is a shorter training time and the advantages of capsular neural network CapsNet include greater precision, reliability and the ability to work effectively with the com-plex tasks. Combination of convolutional neural network CNN and capsular neural network CapsNet as well as improve-ment of CNN-CapsNet architecture were conducted by means of making changes to CapsNet. These changes are that in dynamic routing, the feedback process adds support for the capsule that most closely matches the original signal. Activa-tion functions are also applied to approximate nonlinear connections in deep networks. They are implemented as basic mathematical functions, usually for scalar quantities. Convolutional layers are used to get initial feature maps, which are then loaded into the CapsNet model to perform the final classification. Based on this approach , two separate model have been developed. One model provides classification, based on two classes: \"unripe sunflower\" and \"ripe sunflower\". The second model provides classification based on three classes: \"unripe sunflower\", \"ripe sunflower\" and \"sick sunflower\".Main indicators of the effectiveness of the CNN-CapsNet neural network were ed such characteristics as accuracy, sensitivity and F-score based on type I and II errors. To analyze these indicators, error matrices and graphs of accuracy and errors of these models were constructed. The comparison of the proposed models of the CNN-CapsNet neural net-work with similar ones was made, the highest accuracy was demonstrated by the proposed models.