Improvement of neural network model topology for object segmentation in digital images based on convolutional neural networks