Efficiency of YOLO neural network models applied for object recognition in radar images

Objectives. The paper addresses the problem of applying neural networks for object detection in radar images and their recognition under conditions of limited computational resources. The aim was to investigate the speed and recognition quality of YOLO2 neural network models in solving object detection and classification tasks in radar images in order to evaluate the feasibility of their practical implementation on a microcomputer with a neural processor.Methods. Machine learning, object detection, and classification techniques were used to detect and classify objects in a radar image.Results. The study compared the speed and recognition quality of the 5th, 8th, and 11th generation YOLO neural network models with varying numbers of trainable parameters (nano-, small-, medium-, large-, and extra-largesized) to assess their potential use on a microcomputer with a neural processor. As a result of comparing various YOLO models using evaluation metrics, YOLOv11n (0.925), YOLOv5l (0.889), and YOLOv11s (0.883) showed the highest precision metric; YOLOv5n (0.932), YOLOv11n (0.928), and YOLOv11s (0.914) showed the highest recall metric; YOLOv11s (0.961), YOLOv5n (0.954), and YOLOv11n (0.953) showed the highest mAP50 metric; and YOLOv5n (0.756), YOLOv11s (0.74), and YOLOv5l (0.727) showed the highest mAP50-95 metric.Conclusions. The conducted research confirmed the feasibility of running YOLO neural network models on a microcomputer with a neural processor, provided that the computational resources of the microcomputer match the computational requirements of the neural networks. The ROC-RK3588S-PC microcomputer (Firefly Technology Co., China) provides up to 6 TOPS of performance, allowing the use of YOLOv5n (7.1 GFLOPs), YOLOv11n (6.3 GFLOPs), and YOLOv11s (21.3 GFLOPs) models.