IARet: A Lightweight Multiscale Infrared Aerocraft Recognition Algorithm

Detecting low-altitude unmanned aerial vehicle (UAV) has great significance in both military and civil fields. Aiming at the problem that it is difficult to detect infrared long-range aerocraft targets in complex background accurately and quickly, we propose a lightweight multiscale infrared aerocraft recognition algorithm called IARet. Firstly, in order to improve the detection speed, we use the Focus module at the input to greatly shorten the overall inference time. Secondly, in order to enhance the detection ability, we design a multistage Feature Extraction Part to take into account both lightweight and feature extraction ability, use the Path Aggregation Network to fuse multiscale features, and add an improved Receptive Field Block at the output. Comparison experiments on the Infrared Aerocraft Dataset show that, compared with the newly published high-performance detection model YOLOv5s, the recall of IARet is improved by 2%, the mAP@0.5 is improved by 1.6%, and the CPU inference speed is improved by 37.1%. So IARet model can consider both accuracy and speed. The calculation and parameters of the model are obviously reduced, and the model size is compressed to 4.8 MB, all of which reduce the dependence on the computing capability of hardware platform, and realize the accurate and fast detection of small aerocraft targets in the infrared band.