摘要
CBAM注意力机制可以无缝嵌入神经网络模型中,以提高模型性能。当前国内外学者已在众多智能研究领域引用了CBAM机制。然而,CBAM机制内部池化操作并不能充分表达信号特征。为提高CBAM机制性能,提出利用离散余弦变换频率分量代替池化操作。CBAM机制通道注意力采用各通道前k个最高离散余弦变换频率分量代替池化操作;空间注意力采用通道维度离散余弦变换所有频率分量代替池化操作。通过六组原生和改进CBAM机制对比实验可知:改进CBAM机制和ResNet50的整合后心音分类效果更加精准;前16个最高离散余弦变换频率分量能够充分表达信息通道特征。
The attention mechanism of CBAM can be seamlessly embedded into the neural network model to improve the performance of the model.At present,domestic and foreign scholars have quoted CBAM mechanism in many fields of intelligence research.However,the internal pooling operation of CBAM cannot fully express the signal characteristics.In order to improve the performance of CBAM mechanism,the frequency component of discrete cosine transform(DCT)is proposed to replace the pooling operation.In CBAM channel attention,K maximum DCT frequency components of each channel are used to replace the pooling operation.The spatial attention uses all DCT frequency components of channel dimension instead of pooling operation.Through six groups of experiments comparing the original and improved CBAM mechanism,it can be seen that the improved integration of the improved CBAM mechanism and Resnet50 makes the classification effect of heart sounds more accurate;the first 16 maximum DCT frequency components can fully express the characteristics of the information channel.
作者
张俊飞
张贵英
ZHANG Junfei;ZHANG Guiying(Information and Modern Educational Technology Center,Guangzhou Medical University,Guangzhou 511436,China;School of Basic Medicine,Guangzhou Medical University,Guangzhou 511436,China)
出处
《自动化与仪器仪表》
2021年第9期87-90,95,共5页
Automation & Instrumentation
基金
广东省医学科学技术研究基金(No.A2020194)。
