This paper explores a ship detector from an aerial view, where we focus on predicting ship orientation and improving recall rate for dense objects. First, since ships appear as small objects in remote sensing images, we enhance the backbone network in feature extraction, we reconstruct a classification network as a feature extractor, and then cascade a feature pyramid network for feature fusion. Second, to reduce the adverse effect of post-processing on dense predictions, we use rotated rectangular bounding boxes to represent ships, while adding angle predictions to the dense head. Finally, in order to increase the difference of features at different angles, we propose a target encoding. For the same kind of ships with different orientations, under this encoding rule, other predictions of the network, such as width and height, will have the same regression target. Considering both speed and accuracy, our method performs better in ship detection than other general detectors.


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Arbitrary-Oriented Ship Detection based on Deep Learning


    Beteiligte:
    Chen, Xingyu (Autor:in) / Tang, Chaoying (Autor:in)


    Erscheinungsdatum :

    01.06.2022


    Format / Umfang :

    444231 byte




    Medientyp :

    Aufsatz (Konferenz)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch



    Ship Detection Based on Deep Learning

    Wang, Yuchao / Ning, Xiangyun / Leng, Binghan et al. | British Library Conference Proceedings | 2019


    Research on ship object detection based on deep learning

    Jin, Junkuan / Xiao, Yingjie | SPIE | 2024



    Ship interactions in arbitrary channels

    Hynes, Brodie James, 1974- | DSpace@MIT | 1998

    Freier Zugriff

    A Ship Recognition and Load Detection System Based on Deep Learning

    Li, Honglei / Zhao, Xinlong | Springer Verlag | 2024