ПОСТ-ОБРОБКА ПРОГНОЗІВ ДЛЯ ПОЛІПШЕННЯ ЯКОСТІ РОЗПІЗНАВАННЯ ОБ’ЄКТІВ НА ПОВЕРХНІ ВОДОЙМ

Context. The significance of this work stems from the growing need for UAV technologies integrated with artificial intelligence, aimed at detecting and identifying objects on the surface of water bodies. Modern needs in water body monitoring, especially in the context of environmental monitoring, protection and resource management, require accurate and reliable solutions. This work demonstrates methods for improving the performance of neural networks and offers approaches for processing NN predictions, even if they are trained on irrelevant data, which increases the versatility and efficiency of the technology. Objective. The goal of the work is to solve the problem of false recognition of objects on the surface of water bodies, which is due to a decrease in the accuracy threshold for the neural network. This provides more accurate and reliable detection, reducing the number of false positive predictions and increasing the efficiency of the system in general. Method. It is proposed to add a stage of post-processing of NN predictions, which inherits concepts of min-max suppression used by YOLO models. This algorithm suppresses the re-detection of the object by the network and relies on the cross-sectional area of the detected rectangles. It uses a threshold value of 0.8 for the two points of the rectangle, which can effectively reduce the number of re-predictions and improve the accuracy. Results. As a result of the implementation of the proposed algorithm and the script created on its basis, a result was achieved in which groups from several predictions are combined and filtered. The received data is stored in the database as found and detected objects. The proposed post-processing algorithm effectively removes redundant predictions while maintaining forecast accuracy. This ensures the reliability of the system and increases its performance in real conditions. Conclusions. Detected images of objects on the surface of water bodies are stored in the database in the form of records with unique file name identifiers. After tests with pre-taken images algorithm proved it`s persistence against data duplication scenarios. This increases the efficiency and reliability of the monitoring system, ensuring accurate and timely detection of objects on the surface of water bodies.