The problem of target detection on a radar image (RI) is considered. When solving it, it is proposed to use a finite mixture as a model describing the structure of the processed image. It is assumed that each of its components corresponds to one of the classes of objects present on the radar image, which are similar in their reflective properties. This hypothesis allows one to detect targets by solving the problem of finite mixture separating. It is possible to use for this existing methods of mixture decomposition — the EM-algorithm. However, the structure of the image, in which target classes are local inhomogeneities, consisting of a small number of samples, imposes restrictions on the possibility of using these algorithms in their pure form and leads to the need to create their adaptations that take into account this feature of the input data. The article presents the results of applying the adapted EM-algorithm using the generated radar image with pixels obeying the normal law as an input data. The efficiency of the created algorithm is assessed in comparison with the results of applying the classical version of the EM-algorithm for this model. The data obtained made it possible to reveal the peculiarities of the method due to both the created mechanics of image processing and the properties of the procedure for separating mixtures — the EM-algorithm, which must be taken into account in the further use of this method of image processing.
Достовалова А.М. Применение модели смеси вероятностных распределений в обработке радиолокационных изображений. Математическое моделирование и численные методы, 2020, № 3, с. 117–130.
The article deals with the problem of classifying pixels of the radar image (RI). A locally homogeneous radar image model was used, in which the readings of each small area (local area) were considered to belong to only one class. The classification results of several real radar images by local areas are compared using the statistical criteria for the maximum a posteriori probability, Kolmogorov and Cramer-Mises-Smirnov. At the same time, in the case when the listed criteria made it difficult to classify a local area — when it hit the interface of the underlying surfaces, it was considered to be assigned to a special, boundary class, and its readings were processed using the grid method for separating mixtures of probability distributions. For each criterion, the classification accuracy was evaluated as the proportion of correctly classified pixels within the selected homogeneous areas. It has been established that in the case of significant interclass differences, the best classification accuracy is ensured by the use of the least powerful Kolmogorov criterion among nonparametric criteria. Also, using a real image as an example, it is shown that when the differences in the characteristics of objects of the same class are comparable to interclass differences, the highest classification accuracy is achieved when using the maximum a posteriori probability criterion. Such cases are typical for a wide class of classification problems, including those not related to image processing.
Достовалова А.М. Моделирование локально-однородных радиолокационных изображений при использовании различных статистических критериев. Математическое моделирование и численные методы, 2021, № 4, с. 103–120.