Model of object detection in infrared spectrum images obtained by unmanned aerial vehicles

Unmanned aerial vehicles have found wide application in various fields of monitoring, reconnaissance, remote control. To successfully fulfil these tasks, UAVs are equipped with mobile computer vision systems and computers. Visible images obtained by them may be of insufficient quality due to weather conditions or low illumination. Thus infrared spectrum imagery is the preferred output. This paper presents a neural network model for detecting small objects in infrared spectrum images acquired by UAVs. The model architecture is based on the YOLO5 deep learning model and consists of a basic block and an intermediate block and also includes a prediction block. The basic block is based on the CSPDarknet-53 neural network model and is designed to extract feature maps from the images as input to the model. To describe the intermediate module, it is proposed to use a Bi-FPN neural network that forms a pyramid of feature maps of input images. It was proposed to include coordinate attention modules in the architecture of the Bi-FPN network, which allowed to increase the recognition accuracy while maintaining the computational requirements for mobile machine vision systems. Numerical experiments were conducted on a set of HIT-UAV infrared images, showing the superiority of the proposed model over models such as SSD, Faster RCNN, Retinanet, and YOLO5. Computer experiments showed that the model is able to recognise objects with an accuracy of more than 81.57%.