基于改进Mask R-CNN的卫星目标部位检测方法

针对卫星部件维修更换、燃料加注、废弃卫星回收等空间在轨服务中需解决的目标卫星部位检测问题,在Mask R-CNN的基础上,改进其主干网络结构并缩减分类回归、Mask分支通道数,提出了一种改进的实例分割网络模型Ring-Engine-Mask R-CNN,使用实物模型图像和3dsMax生成的仿真图像建立了专用数据集,给出了一种基于深度学习的卫星目标部位检测方法;实验结果表明,该方法能较好的完成卫星星箭对接环和远地点发动机喷管两种目标部位的检测分割,相较于传统的网络模型,在缩小了模型规模的同时,具有更高精度和更快的检测速度.

Satellite Target Part Detection Method Based on Improved Mask R-CNN

Aiming at the target satellite detection problems that need to be solved in space on-orbit services such as satellite component maintenance and replacement, fuel refueling, and recycling of abandoned satellites, an improved instance segmentation network model Ring-Engine-Mask R-CNN is proposed on the basis of Mask R-CNN by improve the structure of its backbone network and reducing channels of classification, regression, Mask branch. The network uses physical model images and simulation images generated by 3ds Max to establish a dedicated dataset, and presents a satellite target detection method based on deep learning. The experimental results show that this method can complete the detection and segmentation task of the two target parts of the satellite marman ring and the apogee engine nozzle better. Compared with the traditional network model, it has a smaller model volume. Meanwhile, it has a higher detection accuracy and faster speed.