基于多尺度语义的目标检测方法
Multi-scale context extraction network for target detection
曾溢良 1张浩 2吕志武2
作者信息
- 1. 北京科技大学 自动化学院,北京 100083
- 2. 中国航天科工集团第二研究院七○六所,北京 100854
- 折叠
摘要
针对基于卷积神经网络(convolutional neural network,CNN)的检测方法只关注目标的 自身信息,忽略了语义信息,限制 目标检测精度提高的问题,提出一种多尺度语义提取网络,分别提取CNN多层特征图的语义信息并融合,实现目标全局语义和局部语义的提取.在此基础上,将自身特征与语义特征融合,实现目标检测框架中自身特征和语义特征的编码.实验结果表明,该方法与原始的 目标检测网络相比,检测精度有明显提高,尤其是对混叠 目标和小目标具有良好的检测效果.
Abstract
Aiming at the problem that the detection method based on convolutional neural network(CNN)only pays attention to the information of the object itself and ignores the context information,thus limiting the improvement of target detection accuracy,a multi-scale context extraction network was proposed.The context information of CNN multi-layer feature map was extracted and fused separately,and the extraction of global contexts and local contexts was realized.On this basis,the self-features and con-text features were fused to achieve the encoding of self-features and context features in the target detection framework.Experi-mental results show that the detection accuracy of the method is obviously improved compared with that of the original object detection network,especially for overlapping targets and small targets.
关键词
目标检测/深度学习/语义信息/卷积神经网络/多层特征融合/混叠目标/小目标Key words
object detection/deep learning/context/convolution neural network/multi-layer feature fusion/overlapping tar-gets/small targets引用本文复制引用
基金项目
装发预研领域基金项目(6140452010101)
国家自然科学基金项目(61801018)
出版年
2024
NETL
NSTL
万方数据