Weakly supervised arbitrary-oriented ship detection network based on dynamic scale feature enhancement

The advancement of ship detection in remote sensing images has been greatly facilitated by the enhancement of convolutional neural networks. When facing complex marine scenes, oriented ship detection is practical and necessary, whereas the process of developing fully supervised methods is restricted by the annotated data for a rotated box. Therefore, we propose a weakly supervised oriented ship detection network (DSENet) based on dynamic scale feature enhancement, which trains the network with only data annotated in a horizontal box (HBox). First, we propose a dynamic scale feature enhancement module, which can provide precise scale features of the ship for the self-supervised (SS) branch and weakly supervised (WS) branch via feature extraction, fusion, and adaptive enhancement on multiple scales. Next, the neural architecture search is introduced to design the head module to find the optimal structure for the output layers. In addition, a circumscribed Scylla intersection over union loss is designed, which guides the model to predict ship objects accurately through angle-aware vectors and accelerates model training. Finally, DSENet is trained and validated on HRSC2016, dataset for oriented ship recognition (DOSR), and SAR ship detection dataset datasets, respectively. The experimental results indicate that DSENet achieves advanced detection and lightweight performance in WS arbitrary-oriented ship detection.