The paper considers the issues related to the monitoring of the state of steel structures. Such monitoring is proposed to be based on the processing of video images obtained during the inspection of structural elements. The pro-cessing is performed using combinations of artificial neural networks that have passed special training procedures based on transfer learning technologies. This approach allows to overcome the problems of training and post-training of neural networks on small volume samples. The received results allow to claim the possibility of using the developed procedures for searching important classes of defects (in particular, cracks) on images and defining their parameters (size, class, danger degree, etc.). Thus work quality of these procedures (in particular the probability of correct detection and false alarms) appears to be close to the quality of work of competent personnel.