Approaches to the formation of an image database in the explosive ordnance identification system for further training of a neural network

Abstract

The use of unmanned aerial vehicles to detect explosive objects is a promising area of demining. Currently, a significant role in identifying detected objects remains with the operator (human). To solve the problem of identifying images of explosive objects, it is advisable to equip unmanned aerial vehicles or other robotic systems with appropriate software and hardware with built-in elements of artificial intelligence that would allow to recognize detected objects with a certain probability by their characteristic local features (shape/image, body material, body temperature) and classify them without operator participation. Recognition consists in calculating the relevance measure between the descriptions of the object and the reference, whose images are given as sets of descriptor vectors, and optimizing the values of this measure for the reference database. The article discusses the main algorithms (methods) for image identification and selects those that have higher performance and are built according to the binary type, which simplifies the calculation process and reduces costs. In general, the combination of ORB and BRISK descriptors gives better results, since it takes into account color distortion, poor lighting, noise, and different projections. The paper proposes approaches to forming a database of explosive ordnance images, which will further serve as an input layer for training a neural network to identify explosive ordnance. In the course of the work, a program code was written that makes it possible to test the performance of combined methods (algorithms) of image identification in terms of database formation, storage and testing.