A Sea-Surface Target Detection Method based on Weighted Aggregation Network.

Aiming at the problem that the detection performance is easily affected by objective conditions such as light environment and background noise in the complicated ocean environment and there is image scale transformation in congested scenarios and loss of deep semantic information during forwarding, this paper proposed an actual sea surface target detection algorithm combined with the Convolutional Block Attention Module (CBAM) and the Bidirectional Feature Pyramid Network (BiFPN). Firstly, an efficient weighted feature aggregation network was constructed based on the YOLOv5s model by adding the skip connection between input and output nodes, which integrated the deep and shallow semantic information and achieved a unified description of the cross-scale features. On the basis of original feature network, a series of attention feature map information with channel and space two dimensions was generated for meeting the needs of surface target detection for environmental noise suppression and enhancing the feature expression ability of the network in complex backgrounds. And then, for improving the training speed and inference accuracy, the SIoU loss function was introduced as the location loss, which redefined the penalty metrics considering the vector angle between the required regressions to ensure the accurate localization of the prediction box. Finally, the validation was carried out on the actual sea image dataset provided by the Key Laboratory of Marine Intelligent Equipment and System, Ministry of Education. The experimental results showed that the improved model had higher detection accuracy than the YOLOv5s model, the [email protected] :0.95 was increased by 1.8 percentage points, and the speed reached 77FPS. The model showed good generalization performance, and it could meet the requirements of both target accuracy and real-time detection, which could be effectively applied to the ship intelligent perception field.