融合自注意力机制改进ResNet的图像分类方法
Image Classification Method of Improved ResNet by Integrating Self-Attention Mechanism
周录庆 1贾可 2冯翱 1易国锋 3金治成 1李涵鑫 1许昌源1
作者信息
- 1. 成都信息工程大学 计算机学院,四川 成都 610225
- 2. 成都信息工程大学 计算机学院,四川 成都 610225;成都考拉悠然科技有限公司,四川 成都 610000
- 3. 成都考拉悠然科技有限公司,四川 成都 610000
- 折叠
摘要
为解决在大数据集的图像分类任务上,卷积神经网络因缺乏全局信息导致识别准确率受限制的问题,提出将自注意力机制引入卷积神经网络.首先,通过卷积神经网络提取图像特征、改进自注意力模块;其次,基于卷积运算计算注意力的CA模块重构特征图,以突出重要特征并抑制一般特征,为网络加入全局信息;最后,在输出层Avg-pool后引入Dropout层减少过拟合,提升模型鲁棒性和泛化性能.在公开数据集ImageNet-1K、Oxford 102 Flowers和CIFAR-100的实验表明,所提方法识别准确率相较于ResNet50分别提升1.8%、0.72%和13.7%,相较于ResNet50模型的识别性能更优.
Abstract
To solve the problem of limited recognition accuracy in image classification tasks on large datasets due to the lack of global informa-tion in convolutional neural networks,it is proposed to introduce self attention mechanism into convolutional neural networks.Firstly,image features are extracted through convolutional neural networks and the self attention module is improved;Secondly,the CA module based on convolution operation calculates attention to reconstruct feature maps,highlighting important features and suppressing general features,add-ing global information to the network;Finally,a Dropout layer is introduced after the Avgpool output layer to reduce overfitting and improve the robustness and generalization performance of the model.Experiments on publicly available datasets ImageNet-1K,Oxford 102 Flowers,and CIFAR-100 have shown that the proposed method improves recognition accuracy by 1.8%,0.72%,and 13.7%compared to ResNet50,re-spectively;Compared to the ResNet50 model,it has better recognition performance.
关键词
图像分类/卷积神经网络/自注意力机制/卷积运算/DropoutKey words
image classification/convolutional neural network/self-attention mechanism/convolution operation/Dropout引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据