摘要
池化操作是深度卷积神经网络的重要组成部分,也是深度卷积神经网络成功的关键因素之一。然而,在图像识别过程中,传统直接的池化操作会损失特征信息,影响识别的准确率。针对池化操作的特征信息损失问题,提出了基于离散小波变换的双域特征融合模块,以克服直接使用池化操作的缺点。该模块同时考虑了空域和通道域的双域特征融合,将池化操作嵌入在空域特征融合模块与通道域融合模块之间,有效地抑制了直接使用池化操作带来的特征信息损失。通过替换已有的池化操作,新的双域特征融合模块可以非常容易地嵌入到目前流行的深度神经网络架构中。针对图像分类问题,采用VGG,ResNet以及DenseNet等主流网络架构,在CIFAR-10,CIFAR-100,Mini-Imagenet等数据集上进行了一系列实验。实验结果表明,相比经典网络、流行的基于嵌入注意力机制网络和最新基于小波的深度卷积神经网络,所提方法可以获得更高的分类准确率。
Pooling operation is an essential part of deep convolutional neural networks,and also one of the key factors for the success of deep convolutional neural network.However,in the process of image recognition,the traditional direct pooling operation will lead to the loss of feature information and affect the accuracy of recognition.In this paper,a dual-field feature fusion module based on discrete wavelet transform is proposed to overcome the disadvantage of the direct pooling operation.In this module,the dual-field feature fusion of spatial domain and channel domain is considered,and the pooling operation is embedded between spatial feature fusion module and channel feature fusion module,which effectively suppress the information loss of features caused by pooling directly.By replacing the existing pooling operation,the new dual-field feature fusion module can be easily embedded into the current popular deep neural network architectures.Extensive experimental results on CIFAR-10,CIFAR-100 and Mini-Imagenet datasets by using mainstream network architectures such as VGG,ResNet and DenseNet.The experimental results show that compared with the classical network,the popular network based on embedded attention mechanism or latest wavelet basis model,the proposed method can achieve higher classification accuracy.
作者
孙洁琪
李亚峰
张文博
刘鹏辉
SUN Jie-qi;LI Ya-feng;ZHANG Wen-bo;LIU Peng-hui(School of Mathematics and Information Sciences,Baoji University of Arts and Sciences,Baoji,Shaanxi 721013,China;School of Computer,Baoji University of Arts and Sciences,Baoji,Shaanxi 721016,China)
出处
《计算机科学》
CSCD
北大核心
2022年第S01期434-440,共7页
Computer Science
基金
国家自然科学基金(61971005)
陕西省科技厅工业攻关项目(2022GY-064)
宝鸡文理学院研究生创新科研项目(YJSCX21YB09)。
关键词
深度卷积神经网络
池化
注意力机制
离散小波变换
特征融合
Deep convolutional neural networks
Pooling operation
Attention mechanisms
Discrete wavelet transform
Feature fusion
