期刊文献+

多层卷积特征融合的双波段决策级船舶识别 被引量:7

Multi-layer convolutional features fusion for dual-band decision-level ship recognition
下载PDF
导出
摘要 针对可见光和红外双波段船舶识别标注样本少、特征级融合精度低的问题,提出了一种基于多层卷积特征和后验概率加权的决策级融合识别方法。首先,利用预训练卷积神经网络模型,分别提取双波段船舶图像的卷积特征。然后,利用主成分分析方法进行卷积特征降维,设置特征重构阈值自动选择低维空间维度,以适应双波段和各卷积层的特征差异。随后,通过L2范数归一化和级联方法,融合每个波段的中级和高级多层卷积特征。最后,通过加权融合两个波段的支持向量机分类后验概率,构建决策级融合识别模型。实验结果表明:决策级融合识别精度比特征级融合识别精度提升1.5%~2.5%,而且最好值89.7%高出现有最优识别精度1.5%。具有执行简单、处理速度快、识别精度高的优势。 To address the problems of few annotated samples and low accuracy of feature-level fusion in visible and infrared dual-band images for ship recognition,a dual-band decision-level fusion recognition method based on multi-layer convolution feature fusion and posterior probability weighted fusion is pro⁃posed herein.First,convolutional features of dual-band images are extracted from a pre-trained convolu⁃tional neural network model.Subsequently,principal component analysis is employed to reduce the dimen⁃sionality of convolutional features.To satisfy the differences between dual-band features and convolutional layer features,the feature reconstruction threshold is set to select automatically the dimensionality of lowdimensional space.Next,L2 norm normalization and concatenation methods are exploited to fuse mediumand high-level convolutional features for each band image.The classification posterior probabilities of a support vector machine for all bands are then fused by weight factors,and a decision fusion model is finally constructed for dual-band ship recognition.Experimental results indicate that the recognition accuracy of decision-level fusion exceeds that of feature-level fusion by 1.5%~2.5%,and the best mean accuracy reaches 89.7%and is higher than the existing state-of-the-art accuracy by 1.5%.Overall,the proposed method can perform in a simple and rapid manner and achieves high recognition accuracy.
作者 邱晓华 李敏 邓光芒 王利涛 QIU Xiao-hua;LI Min;DENG Guang-mang;WANG Li-tao(College of Operational Support,Rocket Force University of Engineering,Xi'an 710025,China;College of Information Engineering,Engineering University of PAP,Xi'an 710086,China)
出处 《光学精密工程》 EI CAS CSCD 北大核心 2021年第1期183-190,共8页 Optics and Precision Engineering
基金 国家自然科学基金资助项目(No.61102170) 国家社会科学基金资助项目(No.15GJ003-243)。
关键词 目标识别 卷积特征 双波段图像 主成分分析 决策级融合 object recognition convolutional feature dual-band images principal component analysis decision-level fusion
  • 相关文献

参考文献5

二级参考文献20

共引文献82

同被引文献72

引证文献7

二级引证文献21

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部