Automated solution of the problem of detecting industrial objects on a orthomosaic using a neural network

The task of object detection is common for any image type; only the detection quality indicators and the process technological support of differ. The basis for this study was the materials of aerial photography of industrial facilities. Office processing of aerial photography images by overlapping the original photographs to obtain orthophotomaps of the captured area is a complex and costly process, its automation is fragmentary. This article contains a solution for automating the stage of constructing the industrial object contours as a part of the process of their detection on an orthophotomap. As an approach that provided automation, the authors used modeling of a trained convolutional neural network with a one-stage pass through the SSD algorithm and based on the backpropagation method. The result was the development of a software package capable of identifying and classifying several objects on an orthomosaic. For optimal work with the software package, there are described types and methods of storing generated data, as well as the implementation of the process of a transition from an image coordinate system to a spatial coordinate system using orthomosaic mapping files. The practical significance of the results is in the fact that all steps in the development of the software package are described: there is the given rationale for choosing technologies and algorithms, there are identified and formulated requirements for the software package; there is a described process of training a neural network and the project structure. This allows not only reproducing the proposed solution of the automation problem, but also scaling it considering the input parameters for detecting industrial objects on orthomosaics.