2023 , Volume 28, № 6, p.68-80

A brief overview of the main approaches used for solving the problems of detection, determination of coordinates and classification of objects by their images is performed using neural network technologies. The most common YOLO and Faster RCNN methods based on predicting the position of objects are considered using one- and two-stage detectors. At the first stage of the two-stage methods, we identify areas with a high probability containing objects inside themselves by selective search or using a special neural network. At the second stage they are examined by the classifier to determine whether or not they belong to specified classes, as well as to clarify their location and size. One-step methods do not use a separate algorithm to generate regions of interest. Instead, they simultaneously form a certain number of bounding boxes with different parameters in one pass and predict the class of the object. An experimental comparison of the effectiveness of these approaches has been performed. It is noted that single-stage methods allow processing video streams in real time, but they are inferior in accuracy when separating close classes of objects of complex shape.