Home – Home – Jornals – Avtometriya 2025 number 4

The algorithm proposed by the author earlier for searching for small objects against a background of an inhomogeneous random texture has been modified. The modification allows one to work in parallel with multiple image segmentation. A weight is assigned to each segmentation, which is searched for by means of training based on the error backpropagation method. In the implemented process, five weights are determined by a set of a thousand of photographs of clouds measuring 1200×1200 pixels, on which model objects are applied - circles with a radius of two pixels. Initially, all weights are the same and are equal to 0.2. After learning, the maximum and minimum weights turn out to be 0.404 and 0.116. Graphs showing changes in weights indicate that the learning procedure converges. According to the results of the computational experiment, the modification turns out to be more effective than the original algorithm: the number of errors of the first and second kinds decreases by 1.23 and 1.8 times.