摘要
针对现有卷积神经网络结构不能充分挖掘高低级语义特征的缺点,提出一种联合自学习的属性金字塔模块(JSLAPM),其由多维自学习模块(MSLM)和特征金字塔模块(FPM)构成,可应用到任意主干网络结构中。MSLM通过学习空间维度以及通道维度上的特征矩阵重要性来调整特征矩阵,而FPM通过融合不同深度的特征矩阵来增强特征属性的表达能力。此外,结合通道分离提取模块(CSEM),提出了一种特征金字塔注意力网络(FPANet)。实验结果表明,所提出的网络模型可将VGGNet和ResNet的主干网络精度分别提升了近3%。
In view of the shortcomings of the existing convolutional neural network structure that cannot fully excavate high and low-level semantic features,a new joint self-learning attribute pyramid module(JSLAPM)is proposed in this paper,composed of multidimensional self-learning module(MSLM)and feature pyramid modul(FPM),which can be applied to any backbone network structure.MSLM adjusts the feature matrix by learning the importance of feature matrix in spatial dimension and channel dimension,while FPM enhances the expression ability of feature attributes by fusing feature matrix of different depths.In addition,a self-learning feature pyramid attention network(FPANet)is also proposed based on the channel separation and extraction module(CSEM).Experimental results show that the proposed network model can improve the accuracy of VGGNet and ResNet backbone networks by nearly 3%respectively.
作者
孔令军
赵子昂
刘伟光
周耀威
KONG Ling-jun;ZHAO Zi-ang;LIU Wei-guang;ZHOU Yao-wei(Jinling Institute of Technology,Nanjing 211169,China;Nanjing University of Posts and Telecommunications,Nanjing 210094,China)
出处
《金陵科技学院学报》
2021年第3期13-21,共9页
Journal of Jinling Institute of Technology
基金
中国博士后科学基金资助项目(2020M671595)
江苏省博士后科研资助计划资助项目(2020Z198)
金陵科技学院高层次人才科研启动基金(jit-b-202110)。
关键词
卷积神经网络
人体关键点检测
特征金字塔
注意力机制
convolutional neural networks
human keypoint detection
feature pyramid
attention mechanism
