改进的RetinaFace复杂构件关键点检测算法  

A KEYPOINTS DETECTION ALGORITHM FOR COMPLEX COMPONENTS BASED ON IMPROVED RETINAFACE

作　　者：林鑫 沈建新[1] 秦顺 潘峰 Lin Xin;Shen Jianxin;Qin Shun;Pan Feng(College of Mechanical&Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,Jiangsu,China)

机构地区：[1]南京航空航天大学机电学院,江苏南京210016

出　　处：《计算机应用与软件》2024年第8期311-318,共8页Computer Applications and Software

基　　金：国防重大项目培育基金项目(NP2020420)。

摘　　要：为了解决航天装备复杂构件进行喷涂作业时难以定位、缺少关键点数据集等问题,在三维建模软件上搭建构件模型,通过截图和标记关键点制作数据集,并针对数据集量少的问题采取数据增操作。研究现有的RetinaFace关键点检测算法并进行改进,将主干特征提取网络采用优化的MobileNet结构,学习率采用余弦退火衰减,算法输入、输出张量长度与不同构件对应的关键点数相一致。实验结果表明,模型迭代500轮后在验证集上的平均误差降至0.062,能够有效地检测出待喷涂构件的关键点,性能优于同类算法。In order to improve the problems of difficulty in positioning of spraying operation,and lack of datasets in keypoints detection of complex aerospace components,models were built in the 3D modeling software to make datasets of keypoints detection by taking screenshots and marking keypoints.Data augmentation methods were used in order to solve the problem of small sample size of datasets.On the basis of researching and improving existing RetinaFace keypoints detection algorithm,an optimized MobileNet structure was designed for the backbone feature extraction network and the learning rate was decayed by cosine warmup.The length of the input and output tensor was consistent with the number of keypoints corresponding to different components.The experimental results show that the average error on the validation set drops to 0.062 after 500 iterations of the model.The algorithm has better performance than similar algorithms,and can effectively identify the keypoints of components to be sprayed.

关 键 词：航天设备 复杂构件 关键点检测 数据集 RetinaFace 余弦退火衰减 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]