Abstract <p>The efficiency of using a neural network for detecting ‘‘point’’ objects, the shape of which is determined by the characteristics of the recording optical-electronic channel, in images with a spatially nonstationary background is studied. The training data sets take into account the diversity of background situations and the variability of the shape of the signal from objects due to their movement. The results of detecting objects of different brightness and colors on real background images obtained during observation of the Earth’s surface from a geostationary orbit are presented.</p>