基于双通道交叉融合的卷积神经网络图像识别方法研究
Research on the Image Recognition Method of Convolutional Neural Networks Based on Dual-channel Cross-fusion
黄曼曼 1王松林 1周正贵 1侯秀丽1
作者信息
- 1. 安徽商贸职业技术学院 信息与人工智能学院,安徽 芜湖 241002
- 折叠
摘要
针对单通道卷积神经网络的特征提取不够充分、深度网络存在训练困难的问题,提出一种双通道交叉融合的卷积神经网络模型.该模型包括三个特征提取阶段,每个阶段分两条通道进行图像卷积,当两条通道的卷积结束后进行特征交叉融合,经过三次交叉融合后输入到全局平均池化层以及全连接层中得到分类结果.将该模型应用于Cifar10、Cifar100 和Fashion-MNIST的图像分类任务以验证模型的有效性.结果表明,双通道交叉融合模型可以在当前支持GPU加速的主流笔记本电脑上进行训练,在同样规模的数据集上具有比同类其他模型更好的分类性能.
Abstract
In response to the problems of insufficient feature extraction in single-channel Convolutional Neural Networks and the difficulty in training deep networks,a dual-channel cross-fusion Convolutional Neural Networks model is proposed.This model includes three feature extraction stages,with each stage performing image convolution through two separate channels.After the convolution of the two channels is completed,feature cross-fusion is carried out.After three rounds of cross-fusion,the features are input into a global average pooling layer and a fully connected layer to obtain classification results.This model is applied to image classification tasks on Cifar10,Cifar100,and Fashion-MNIST to verify its effectiveness.The results show that the dual-channel cross-fusion model can be trained on current mainstream laptops that support GPU acceleration,and it exhibits better classification performance than other similar models on datasets of the same scale.
关键词
卷积神经网络/双通道/融合/分类/准确率Key words
Convolutional Neural Networks/dual-channel/fusion/classification/accuracy引用本文复制引用
基金项目
安徽省质量工程项目(2021zyyh021)
安徽省质量工程项目(2021jyxm0461)
安徽商贸职业技术学院重点科研项目(2021KZZ01)
安徽省教学示范课建设项目(2020SJ1069)
安徽省精品课程项目(2022jpkc048)
出版年
2024
国家科技期刊平台
NSTL
万方数据