Attention-based feature pyramid networks for ship detection of optical remote sensing image

Ship detection on spaceborne optical images is a challenging task that has attracted increasing attention because of its potential applications in many fields. Although some ship detection methods have been proposed in recent years, many obstacles still exist because of the large-scale and high complexity of optical remote sensing images. Identifying ships from interferences, such as the features of clouds, waves, and some land architectures that are similar to ships, is difficult. Therefore, an accurate and stable deep-learning based method is proposed in this work. The method involves three steps: First, the image feature pyramid is extracted using convolution to detect multiscale ship targets. Second, a multilevel attention feature mapping structure is constructed from top to bottom using the fine-grained features of the top layer from the pyramid to improve the expressive ability of shallow features. Finally, Softmax classifier is used for multilevel ship detection. The experimental results based on real remote sensing images are shot by JL-1 satellite, Google satellite, and NWPU VHR-10. The result proves that the performance of our algorithm is better than the three other state-of-the-art methods. In addition, the network was cut while ensuring accuracy. The complexity of our algorithm is reduced, and its practicality is improved by experiments and analysis. This work proposes an attention-based method called A-FPN. However, unlike traditional algorithms, A-FPN has higher robustness and wider range of use. Furthermore, we effectively cut the network to reduce the complexity of the algorithm, thereby exhibiting the significance of our algorithm in practical applications. © 2020, Science Press. All right reserved.