S&R-Net: An SAR Image Segmentation and Retrieval Network for Ship Database Building

Problems such as background interference and various sizes and types of marine ships still exist in the field of SAR ship image segmentation and retrieval. In this paper, we present a deep learning network for SAR image segmentation and ship retrieval named S&R-Network, which includes a feature encoding model, an image segmentation model and an image retrieval model. In the feature encoding model of the network, feature extractors are designed to obtain a spatial importance map (SIM) and guarantee feature interaction. To further increase the spatial interaction of the feature map and improve feature extraction for specific objects, RotationGate, which projects input features before the process of CoordGate, resulting in feature encoding for specific objects, is designed. In addition, we decode the feature map in the image segmentation model, which removes SAR image background interference and prepares for SAR ship retrieval. In the image retrieval model, we design a feature memory bank to build a database, which allows the network to output the highest similarity to the retrieved image by calculation. Simulation experiments conducted on the annotated OpenSARShip dataset demonstrate that our proposed network outperforms other networks in terms of SAR image segmentation. Moreover, we achieve good performance in SAR ship retrieval, which proves the applicability of the proposed network. Compared with several classical image processing networks, our method reached 87.21% mIoU, 96.14% mSA, 96.6% precision and 94.53% recall in segmentation and 88.15% RAinthe image retrieval task.