High-accuracy and real-time satellite component semantic segmentation can locate the key satellite components, such as solar panels, to be operated in on-orbit services, which is of great significance for navigation and control. However, to accomplish the above aim, two main challenges remain unsolved. First, satellite component semantic segmentation algorithms require a large number of images for training; however, on-orbit satellite images are difficult to obtain, especially for a large-scale satellite component video dataset. In addition, high-accuracy semantic segmentation networks require relatively more computation resources, which are difficult to be fulfilled in on-orbit tasks. How to build a satellite component semantic segmentation network that meets the requirements of both high-accuracy and real-time on-orbit operation is the key aim to be accomplished in this article. In this article, a simulated satellite component dataset consisting of 98 video sequences of 13 satellites, with a complex background, various on-orbit illumination, and common satellite motion, is proposed, and it has 32 402 frames in total. To meet the requirements of both high-accuracy and real-time on-orbit operation, this article proposes an attention-based real-time network, Pyramid Attention and Decoupled Attention Network (PADAN), which contains an image-based version, PADAN-S, and a video-based version, PADAN-T. The PADAN-S, which mainly adopts pyramid attention calculation on three-layer pyramid features and then performs decoupled attention calculation by considering both row and column attention, is based on AttaNet. The PADAN-T uses a part of the PADAN-S to obtain temporal pyramid features from temporal frames, then performs decoupled attention calculations between the features of the output frame and the features at each layer in the temporal pyramid. The experimental results show that the PADAN-S and PADAN-T have superior performance compared to other real-time state-of-the-art algorithms in accuracy in both image-based and video-based satellite component semantic segmentation tasks on simulation datasets, and our dataset has a degree of simulating the real on-orbit environment. The PADAN-S can achieve a speed of 10.25 frames per second with an image solution of 1280 pixels × 720 pixels on the edge computing device Jetson Xavier, and the PADAN-T can obtain a speed of 7.18 frames per second.


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Satellite Component Semantic Segmentation: Video Dataset and Real-Time Pyramid Attention and Decoupled Attention Network


    Beteiligte:
    Shao, Yadong (Autor:in) / Wu, Aodi (Autor:in) / Li, Shengyang (Autor:in) / Shu, Leizheng (Autor:in) / Wan, Xue (Autor:in) / Shao, Yuanbin (Autor:in) / Huo, Junyan (Autor:in)


    Erscheinungsdatum :

    2023-12-01


    Format / Umfang :

    8290058 byte




    Medientyp :

    Aufsatz (Zeitschrift)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch




    RGB-D Co-attention Network for Semantic Segmentation

    Zhou, Hao / Qi, Lu / Wan, Zhaoliang et al. | British Library Conference Proceedings | 2021



    A hybrid attention semantic segmentation network for unstructured terrain on Mars

    Liu, Haiqiang / Yao, Meibao / Xiao, Xueming et al. | Elsevier | 2022


    Efficient Semantic Segmentation via Self-Attention and Self-Distillation

    An, Shumin / Liao, Qingmin / Lu, Zongqing et al. | IEEE | 2022