An approach is proposed to substantiate the architecture of a convolutional neural network (SNS), which provides improved performance in the number of weighting factors and connections while maintaining the accuracy of object recognition in images.
Learning SNA with the selected architecture is implemented on Earth. After training, the neural network is used in onboard computing systems for autonomous recognition of objects in images. The analysis of the features of the construction of the SNA for the recognition of objects in images is carried out, and on its basis, limitations and rules for the development of new SNA are formulated. The formulation of the combinatorial optimization problem is formulated, a heuristic algorithm is proposed for solving this problem based on the rules of layout and cutting off unpromising SNA configurations. For the Planetsnet and MNIST test datasets, the results of computer experiments are presented and compared with existing SNA architectures by the number of weights, connections, and accuracy in% of errors on the test set. It is shown that the proposed algorithm allows you to choose a variant of the SNA architecture with a smaller number of weighting factors and compounds with almost the same accuracy in% error solving the problem of recognizing objects in the image.