Multiframe Detection of Sea-Surface Small Target Using Improved YOLOv5

Sea-surface small target detection in sea clutter background has been a challenging problem for maritime surveillance radar. The dynamic sea surface usually results in a major loss and false alarms in target detection. Therefore, this paper proposes a multiframe detector of sea-surface small target using the improved YOLOv5, to reduce false and missed alarms. Specifically, in order to increase receptive fields for better feature extraction, the Receptive Field Block (RFB) is introduced into the network. To increase the feature information for detecting small target scales and fuse more features without increasing the cost, the Bidirectional Feature Pyramid Network (BiFPN) is added in the neck module. Moreover, the Simple and the Parameter-Free Attention Module (SimAM) is used to pay more attention to small object features. The proposed approach has been verified on the actual radar measurement data of the Yantai port. The results demonstrate the effectiveness of the proposed algorithm in sea-surface small target detection.