Recognition and Instance Segmentation of Space Non-cooperative Satellite Components Based on Deep Learning

Abstract

With the rapid development of aerospace technology, the future aerospace on-orbit servicing and control tasks will face particularly challenging problem of more unknown and changeable operation targets and more diversified and refined operation tasks. Accurate identification and pinpointing the non-cooperative satellite key components such as the inspection and tracking, flexible capture and on-orbit maintenance are important prerequisites and safeguards for on- orbit servicing and control tasks. Aiming at the problem that the recognition and segmentation of space target satellite components are faced with various components, large structural differences and uneven scale. A synchronous recognition and instance segmentation algorithm based on improved Mask RCNN for space non-cooperative satellite components is proposed in this paper. The image features are extracted by the ResNet101 backbone network, and then divide the post-processing feature map into two parallel task branches: detection branch and segmentation branch. The detection branch includes classification and box regression, which is completed by using region-based Faster RCNN. The segmentation branch uses the HRNetV2p network to fuse the high, medium and low resolution convolution parallel. The edge information is better preserved and the prediction of large and small objects mask are both taken into account with alterable resolution. Considering the particularity of space tasks, the multi-task loss function is trained based on the satellite physical model sample data sets to obtain the global optimal solution by learning continuously and improve the performance of the network model. In order to verify the feasibility and effectiveness of the algorithm proposed ,the synchronous recognition and segmentation experiments test for the 2° /s around Z-axis real-time motion satellite physical model components at different distances from far to near are completed under the ground semi-physical experimental system to verify the online application ability of the algorithm. The improved Mask RCNN is compared with the original Mask RCNN and the contrast experiment is carried out in the typical aerospace mission application scenario. The experimental results show the improved Mask RCNN can improve the speed of the algorithm while ensuring the accuracy of recognition and segmentation. The satellite components average recognition accuracy and mean intersection-of-union were 2.4% and 7.2% higher than the Mask RCNN model respectively. The average recognition speed shortened by 2.25 ms /frame. Which it is of great significance to the intelligent development of the space on-orbit servicing and manipulating.