Cognitive radar, through intelligent and adaptive processing, enables automatic analysis and decision-making capabilities within radar systems. One of the key tools for achieving this goal is artificial intelligence (AI), particularly deep learning. As a branch of AI, deep learning can play a critical role in radar target detection, especially in noisy and complex environments. In this study, two deep learning networks, Faster R-CNN and YOLOv8n, are employed for radar target detection on simulated data. The performance of these networks is eva‎luated on datasets containing both white and colored clutter. Due to data limitations, training and eva‎luation are conducted using simulated datasets. Additionally, transfer learning is used to enhance the performance of these networks. Transfer learning involves using pre-trained models on one dataset to accelerate and improve learning on a new dataset. The performance of these networks is compared with classical detection methods, namely CA-CFAR and OS-CFAR. These traditional radar target detection approaches face limitations in detection accuracy under noisy conditions and low Signal-to-Noise Ratio (SNR). The results of this study demonstrate that Faster R-CNN and YOLOv8n outperform CA-CFAR and OS-CFAR at low Signal-to-Clutter Ratios (SCR), exhibiting a superior ability to detect targets with a low false alarm probability. For instance, under colored clutter conditions and SCR values below -10 dB, while the accuracy of conventional methods declines, the detection probability of both networks exceeds 90% at a false alarm probability of 1.5 × 10^(-6). These findings underscore the potential of deep learning to improve the efficiency of radar systems and act as a replacement or complement to classical methods. Furthermore, to investigate the performance of these networks in multi-target scenarios, a study is conducted on their resolution capabilities. It is shown that when the distance between two targets exceeds 2 cells, the probability of missing one of the targets is less than 30% for both networks. In conclusion, this research highlights the importance of employing deep learning tools in radar systems and emphasizes the need for further research and development in this domain.

محمدرضا تابان

محمود مدرس هاشمي , محمدرضا احمدزاده