Advances in deep learning have led to its widespread use in UAV ship inspection. However, achieving and maintaining superior performance in complex nearshore environments remains a difficult task for UAV ship detection, taking into account model size and real-time requirements. After weighing the pros and cons between detection accuracy and the number of parameters as well as other performance metrics, we propose CSC-Net, a new network architecture built on YOLOv11. In the inference phase we introduce the wide diverse branch block (WDBB) to significantly enhance the feature extraction capability of the network without affecting the computational effort by reparameterizing the structure; in the Neck part we innovatively propose a contextual feature guided fusion module (CGFM) to amplify the role of important features while downplaying the role of minor features. Meanwhile, to offset the increase in computation due to the accuracy improvement, we also design a lightweight shared detail enhancements convolutional detection head (LED). LED significantly improves the model's capture and generalization capabilities while keeping the computational resource consumption and inference time low. We carried out experiments on three commonly used ship datasets for remote sensing, namely HRSC2016, ShipRSIagmeNet, and DOTAₛhip as well as two commonly used ship datasets SeaShips, Shipsdataset. The experimental results demonstrate that CSC-Net exhibits excellent detection performance compared with other leading UAV ships detection methods.
Miao et al. (Sat,) studied this question.