This paper demonstrates the effective capabilities of deep neural networks in solution of the problem of structural identification on graphs in conditions of a priori uncertainty, incomplete observability and stochastic disturbances which is also knows as subgraph detection or recovery. The problem of identification of observed constellations in a photo of the night sky was considered as a test. The solution with quality of 0.927 𝐹1 is obtained. In this work we synthesized original ResNet architecture of the convolution neural network with 26 trainable layers, 415 193 configurable parameters, carried out statistical analysis of structural characteristics of the dataset and adapted the standard binary cross entropy loss function, developed a special strategy for learning the neural network. Moreover, an adequate criterion of observability of the constellation in the image was formed. We also studied the influence of noise on the quality and stability of the received solutions.