基于鲁棒主成分分析的运载火箭焊缝射线数字成像分类方法  

A Classification Method for Rocket Weld Seam Radiographic Digital Imaging Based on Robust Principal Component Analysis

作　　者：王宁 刘晓 刘骁佳 危荃[1] WANG Ning;LIU Xiao;LIU Xiaojia;WEI Quan(Shanghai Spaceflight Precision Machinery Institute,Shanghai 201699,China)

机构地区：[1]上海航天精密机械研究所,上海201699

出　　处：《上海航天(中英文)》2025年第1期149-156,共8页Aerospace Shanghai(Chinese&English)

基　　金：国家自然科学基金资助项目(92267201)。

摘　　要：运载火箭焊缝射线数字图像自动检测技术主要是对运载火箭焊缝射线数字图像进行分类,而实际生产过程中获取的图像数量庞大,对全部图像进行标注会浪费大量的人力和物力。针对先验的监督信息能够提高目标提取的精度,以及去除图片中的背景可以提高分类精度的问题,本文提出了基于鲁棒主成分分析(RPCA)的拉普拉斯特征映射(LE)正则化半监督目标特征提取算法(SSRLE)。SSRLE以RPCA为基础,在保证数据全局结构的基础上,通过加入自适应邻域图权重矩阵LE正则化保证数据的局部结构,并排除了经典LE算法中近邻值k的影响。在先验信息的作用下,该方法可以很好地分离目标与背景。利用目标数据与监督信息训练线性分类器,并结合与流形平滑假设实现对无标记数据的预测,从而达到较好的分类效果。最后,本文通过实验验证了所提出的算法的有效性,比较了不同半监督算法的分类效果,证明了本文所提方法优于其他方法。The automatic detection technology for digital radiographic images of launch vehicle welds primarily involves the classification of digital radiographic images of launch vehicle welds.However,the actual production process yields an enormous volume of images,and annotating the entire dataset would entail a considerable waste of manpower and resources.Considering that prior supervisory information can enhance the precision of target extraction and that removing the background from images can improve classification accuracy,this paper proposes a Semisupervised Target Feature Extraction algorithm with Laplacian Eigenmaps(LE)Regularization based on Robust Principal Component Analysis(RPCA),termed SSRLE.On the premise of ensuring the global structure of the data,the local structure of the data is guaranteed by adding the LE regularization of the weight matrix of the adaptive neighborhood graphs,and the influence of the nearest neighbor value k in the classical LE algorithm is excluded.Under the influence of prior information,the target and background are separated effectively.The linear classifier is trained with the target data and supervision information.With the manifold smoothing hypothesis,the trained linear classifier can predict unlabeled data,resulting in improved classification results.Finally,experiments are carried out,and the classification effects of different semi-supervised algorithms are compared.The results show that the proposed method is valid,and is superior to other methods.

关 键 词：运载火箭贮箱 鲁棒主成分分析 视觉目标特征提取 流形学习 半监督学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]