基于正则化正交非负矩阵分解的旋转目标检测方法  

Rotating Target Detection Method Based on Regularized Orthogonal Nonnegative Matrix Factorization

作　　者：谢余庆 黄旭东 胡丽莹 XIE Yuqing;HUANG Xudong;HU Liying(College of Computer and Cyber Security,Digital Fujian Internet-of-Things Laboratory of Environmental Monitoring,Fujian Normal University,Fuzhou 350117,China;Weiku(Xiamen)Information Technology Company Limited,Xiamen 361008,China)

机构地区：[1]福建师范大学计算机与网络空间安全学院,福建师范大学数字福建环境监测物联网实验室,福建福州350117 [2]维库(厦门)信息技术有限公司,福建厦门361008

出　　处：《福建师范大学学报（自然科学版）》2024年第1期106-115,共10页Journal of Fujian Normal University：Natural Science Edition

基　　金：国家自然科学基金资助项目(U1805263)。

摘　　要：小样本的旋转目标检测是指在样本数少的情况下进行旋转目标检测模型的训练,深度学习在旋转目标检测领域往往需要庞大的样本数和计算算力。现有的基于机器学习的旋转目标检测方法大多有着对目标尺度和姿态敏感的缺点。因此提出一种基于正则化正交非负矩阵分解的旋转目标检测方法,来解决小样本的旋转目标检测难题。首先,针对样本不具有各种角度的图片,对样本进行旋转后进行背景填充,这样便于更好的表征学习。其次,提出一种基于正则化正交非负矩阵分解算法对旋转样本的梯度直方图特征进行表征学习。最后,为了测试算法在特征学习后的有效性,利用支持向量机对特征提取后的数据进行训练和测试。实验结果表明本文的目标检测方法在多个数据集中可以取得不错的效果。Rotating target detection with small samples refers to training a rotating target detection model under the condition of few samples.In the field of rotating target detection,deep learning often requires a large number of samples and substantial computational power.Most existing machine learning-based rotating target detection methods are sensitive to target scale and posture.To address the challenge of rotating target detection in small samples,this paper proposes a rotating target detection method based on regularized orthogonal nonnegative matrix factorization.Firstly,for samples lacking images at various angles,rotation and background filling are applied to facilitate better representation learning.Secondly,an algorithm based on regularized orthogonal nonnegative matrix factorization is introduced to perform representation learning on the gradient histogram features of rotating samples.Finally,in order to evaluate the effectiveness of this algorithm after feature learning,support vector machines are used to train and test the data after feature extraction.The experimental results show that the target detection method proposed in this paper performs well across multiple datasets.

关 键 词：正则化 正交非负矩阵分解 梯度直方图特征 旋转目标检测 支持向量机 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]