FAST反射面无靶标测量节点识别方法的研究

FAST is the world's largest single-aperture radio telescope built by China. The 2225 nodes on the reflector are the control points and are therefore the objects of the antenna surface measurement. The problems of long measuring time and difficult target maintenance exist in the current laser measuring system of reflector surface. Regarding this, a non-target photogrammetric method is proposed, which uses the splicing gap between panels as the basis of feature recognition and accurately identifies and calculates the position of each node. Therefore, the identification of nodes by using gaps is the foundation and key technology. The purpose of this project is to identify the nodes in the photographs through in-depth graphics and Image processing, and to form a stable and reliable automatic recognition system. After a preliminary study, a feature intersection extraction algorithm based on dynamic threshold and image rotation and a convolutional neural network (CNN) method using LeNet-5 model are proposed. The experiments prove that both methods have good robustness and generalization ability.The recognition rate of both of them can reach more than 90%.It is a key point for the future application of non-target measurement of FAST reflector.

FAST是我国建造的世界最大单口径射电望远镜，其反射面上的2225个节点是反射面的控制点，也是望远镜面型检测的目标点。目前使用的反射面激光测量系统存在测量时间长和后期靶标维护难的问题。对此提出一种无靶标摄影测量方法，利用面板之间的拼接缝隙作为特征识别基础，精确识别计算出各节点位置。因此，利用缝隙进行节点的识别是其基础和关键技术。本项目旨在通过图形深入处理识别出摄影相片中的节点，形成稳定可靠的自动化识别系统。经过初步研究，提出一种基于动态阈值和图像旋转的特征交叉点提取算法和使用LeNet-5模型的卷积神经网络方法。实验证明，两种方法都具有较好的鲁棒性和泛化能力，识别率均可达90%以上，为将来无靶标测量在FAST反射面的实际应用奠定了坚实的基础。

对于FAST射电望远镜而言，其主动反射面由三角形单元面板通过节点连接组成，在获取的图像中节点的图像特征相对于反射面其他区域是显著的，因此可以另辟蹊径寻找图像中的特征信息。我们分析了FAST主动反射面的构成情况，可以通过其自然特征识别到反射面的控制节点，而不需要合作靶标。本项目使用了一种基于动态阈值和图像旋转的特征交叉点提取算法和LeNet-5模型的卷积神经网络方法来进行节点的识别，并对高精度测量相机进行了适应性改造。形成了稳定可靠的识别系统，具有较好的鲁棒性和泛化能力，识别率达91%以上，为将来无靶标测量在FAST反射面的实际应用奠定了坚实的基础。