A Yolo-like lightweight ship detection network for unmanned surface vehicles

Abstract The widespread application of unmanned surface vehicles (USVs) in maritime surveillance has highlighted the need for improved ship detection models. However, hardware limitations and environmental interference impact the perception capability of USVs. In order to address these challenges, YoloS, a lightweight ship detection network, is specifically designed for USVs in complex backgrounds. Firstly, a network component Split Widely Network (SWNet) is put forward to the backbone to reduce its redundancy. SWNet leverages inter-channel interaction and the idea of group convolution to minimize redundancy of the network. Secondly, the Small Pyramid Network (SPN) is introduced as the neck network. SPN enhances the spatial information of the target regions by capturing and highlighting significant contour and texture details within target regions. SPN also utilizes these low-level features with enhanced spatial information to guide high-level features in identifying the most discriminative fine-grained details within the target regions. In order to further achieve network lightweighting, only two output heads for low-level features are retained while still preserving the capability to extract task-critical features. Extensive experiments on different datasets have verified the effectiveness of the proposed method, and it shows that YoloS can achieve 96.5$\%$ detection accuracy and 80.1 fps on the SMD dataset with only 2.78M model parameters and 10.4G floating point operations.