NEURAL NETWORK SYSTEM FOR SELECTING INDIVIDUAL OBJECTS ON RAST IMAGES

Abstract

The article is devoted to solving the problem of increasing the efficiency of neural network tools for semantic segmentation of images. Based on the results of the analysis, it is shown that one of the areas of improvement of such tools is the development of the architecture of the neural network system for the selection of individual objects on raster images. As a result of the conducted research, the architecture of the neural network system for the selection of objects on raster images has been developed, which, due to the adaptation of architectural parameters to the features of the construction and use of modern neural network models intended for the semantic segmentation of images, ensures sufficient accuracy with the permissible amount of use of computing resources. The difference of the developed architecture is the use of functional blocks that are related to the formation of training databases, training of a neural network and selection of an object in the image with the help of a trained neural network. The results of the conducted experiments showed that the application of the proposed architectural solutions allows to develop tools that ensure the achievement of image segmentation accuracy of about 0.8, which corresponds to the accuracy of the best known systems of similar purpose. It is shown that the further increase in accuracy, which can be realized by modifying the parameters of convolutional neural networks on which the encoder and decoder are based, requires additional theoretical research. In addition, the perspective of research related to the improvement of neural network models in the direction of their adaptation to the selection of objects in the video stream is shown