RECOGNITION OF EMOTIONS BY FACIAL GEOMETRY USING A CAPSULE NEURAL NETWORK

The article is devoted to the problem of improving the efficiency of neural network means of emotion recognition by the geometry of the human face. It is shown that one of the most significant drawbacks of modern neural network means of emotion recognition, which are used in General-purpose information systems, is the lack of recognition accuracy under the influence of characteristic interference. It is proposed to improve the accuracy of recognition through the use of capsule neural network model, which has increased adaptability to the analysis of noisy images. As a result of the research, a neural network model of the CapsNet type was developed, designed to recognize basic emotions taking into account such interference as face rotation. It is shown experimentally that in the analysis of undistorted images CapsNet slightly exceeds the accuracy of the classical convolutional neural network type LaNet, which is approximately equal to its resource intensity. The accuracy of CapsNet recognition of undistorted images is somewhat inferior to modern types of convolution networks, which have a much higher resource consumption compared to it. When detecting emotions on rotated images, the accuracy of CapsNet is comparable with the accuracy of modern types of convolution networks and significantly exceeds the accuracy of LaNet. Prospects for further research in the field of neural network recognition of emotions on the geometry of the face can be associated with the improvement of architectural solutions of the capsule neural network in the direction of reducing the number of training iterations while ensuring acceptable recognition accuracy