The recent proposed deep learning-based arbitrary-oriented objects detection algorithms increase extra computation burden and could not work efficiently. A one-stage model based on object centers detection is proposed for arbitrary-oriented ship detection. As the centers of objects are free from the influence their distribution directions, the key of the model is to regress the parameters of object's oriented bounding box on the basis of center detection. Firstly, a feature extracting network is designed to achieve feature map and a new feature fusion method is proposed which aggregates the low-level features rich in detailing information and high-level features rich in semantic information together. Then the feature map is entered to three detection branches, which predict of centers, offsets of centers, and size and direction of the oriented bounding boxes respectively. A combined loss function is proposed for the training of the network, and a modified non-maximum suppression algorithm is proposed for removing invalid oriented bounding boxes. The proposed model achieves state-of-art performance in public SAR ship detection dataset with mean average precision as 0.906, outstanding than other methods both in speed and precision.