For satellite interactions missions such as autonomous docking, the key to successful completion of these missions is to autonomously and accurately detect targets’ information. The missions with high precision requirement need not only to obtain the categories and location of target satellite components but also to obtain the high-level information such as the corner point. In this paper, a satellite components detection (SCD) method based on improved CenterMask is proposed, which is the state-of-the-art instance segmentation network, to autonomously and accurately detect the target components. First, the fully convolutional one-stage object detection detector is optimized to better obtain categories and bounding boxes of targets. Next, a spatial-channel attention module is proposed, and it is introduced in the mask branch to improve the performance of segmentation. Finally, a satellite components dataset is built for model training, and the optimized SCD model is obtained after training. To better satisfy missions with a higher speed requirement, a speed-up method is also proposed, which greatly improves the detection speed at the expense of very little accuracy. Experiments show that, compared with CenterMask, our method improves 2.8 and 1.5% of box average precision () and mask , and the speed-up version improves the speed of 5.2 frames per second.
Satellite Components Detection from Optical Images Based on Instance Segmentation Networks
Journal of Aerospace Information Systems ; 18 , 6 ; 355-365
2021-06-01
Article (Journal)
Electronic Resource
English
Ship Detection from Satellite Imagery Using RetinaNet with Instance Segmentation
Springer Verlag | 2023
|Rock Instance Segmentation from Synthetic Images for Planetary Exploration Missions
German Aerospace Center (DLR) | 2021
|Learning Stixel-based Instance Segmentation
IEEE | 2021
|