A Deep Learning-Based Time-Frequency Scheme for Ship Detection Using HFSWR

Compact High frequency surface wave radar (HFSWR) has been widely used in remote sensing of oceanic dynamics and ship targets due to its convenient deployment and low cost. However, when using a constant false alarm rate (CFAR) detector, these systems experience performance degradation primarily because of echo nonstationarity. To address this challenge, a deep learning (DL)-based scheme tailored for identifying ship targets in the time-frequency (TF) domain is presented. To ensure high-quality model training, we develop a semiautomatic annotation approach that uses automatic identification system (AIS) information as a reference and collect a TF dataset named HFSWR-TFD. In addition, inspired by the dynamic snake convolution and triplet attention mechanism, an improved YOLOv5s model named DS-YOLOv5s is designed to effectively capture target ridges. The inference results are filtered using a confidence threshold and then transformed into the range-Doppler domain for final target identification. Experimental results on the newly collected dataset show significant improvements are achieved by DS-YOLOv5s. Compared to its baseline, the DS-YOLOv5s can increase the F1 score by 15.3%, and AP75 by 6.3%. Then, this pretrained DL model is integrated into the entire scheme to make comparison with existing CFAR detectors. With the AIS records as ground truth, our scheme achieves a match rate that is 2.27 similar to 8.17% greater than its CFAR counterparts. Moreover, the quantitative results of the associate tracks further confirm the superiority of the proposed method. In conclusion, the proposed scheme provides an effective and efficient solution for HFSWR ship detection.