Ансамбль дрібних згорткових нейронних мереж для класифікації статі людини у відеопотоці

Abstract

Subjects of the research are neural network models for a person’s gender classification by the image of a person when processing a video stream. The goalis to investigate the effectiveness of individual shallow neural networks and ensembles created from them to solve the problem of classifying a person’s gender in a video stream, which is processed as a sequence of individual frames. Tasks include the development of mathematical models to process a sequence of frames with accumulation using different strategies, investigation of their effectiveness for solving the classification problem, compiling ensembles of shallow convolutional neural networks. Following method sare used: neural networksmodeling, data mining, mathematical statistics, functional analysis, computer modeling. Results follows: it is shown that the classification accuracy can be improved both through the use ofdifferent voting models of the individual framesclassification results, and through the use of ensembles of shallow convolutional neural networks. The insignificant hardware and software resources that are required for their training and use make it possible to increase the classification speed by several times in comparison with the results of classification by neural networks, that havemore complex architecture. Conclusions. The contribution is in the creation of ensembles of shallow neural networks, the general decision in which is made after the generalizationby various voting methods with confidence both the classification results of individual frames and the classification results of the same frame by different networks, which makes it possible to increase the accuracy and speed of classification. The practical significance of the work is in the creation of a method that makes it possible to provide an acceptable classification accuracy and significantly improve performance by using shallow neural network architectures.