A Content-Based Generator Method for Vessel Detection

With the advancement of satellite imaging technology, the interpretation of remote sensing (RS) images has become an important subject. Especially in the object recognition field, accurate acquisition of vessel location and classification information is crucial to develop strategic plans. However, the lack of ship samples in RS images has been hindering the research of ship fine categorization. In this paper, a vessel generation method based on generative adversarial network is proposed to solve the insufficient samples in RS images. Dealing with sample insufficiency by prior global-local segmentation, category image generation and domain translation composition. Experiments on the HRSC2016 dataset show that the generated pseudo-images are highly similar to the real vessel, which verifies the effectiveness of the method. Besides, we constructed a ship dataset containing 10,000 images, which have great significance in vessel classification and localization.