| 基于多分支网络的图像分类算法 |
| |
| 引用本文： | 杨鑫,杨晶东.基于多分支网络的图像分类算法[J].教育技术导刊,2019,18(7):56-59. |
| |
| 作者姓名： | 杨鑫 杨晶东 |
| |
| 作者单位： | 上海理工大学 光电信息与计算机工程学院，上海 200093 |
| |
| 摘 要： | 为提高卷积神经网络在图像分类中的泛化性，提出基于多分支深度神经网络结构。使用ResNet（残差网络）的跨层连接结构构造多分支网络，各分支网络共享中浅层特征提取，深层网络使用不同卷积核尺寸。分别使用独立损失函数产生多梯度对中浅层特征权值进行同步调整。与ResNet的单重网络进行对比实验，结果表明，在具有相同收敛性的前提下，各个分支网络的泛化性都得到一定提高，在多类别数据集中表现出更优性能。
|
| 关 键 词： | 残差网络 多分支网络 泛化性能 |
| 收稿时间： | 2018-11-12 |
Image Recognition Algorithm Based on Multi-branch Network |
| |
| YANG Xin,YANG Jing-dong.Image Recognition Algorithm Based on Multi-branch Network[J].Introduction of Educational Technology,2019,18(7):56-59. |
| |
| Authors: | YANG Xin YANG Jing-dong |
| |
| Institution: | School of Optoelectronic Information and Computer Engineering， University of Shanghai for Science and Technology， Shanghai 200093， China |
| |
| Abstract: | Aiming at improve the convergence and generalization of convolutional neural networks in image classification， we propose a deep neural network structure based on multi-branch network. Using a cross-layer connection structure of ResNet （residual network）， a multi-branch network is constructed， shallow feature extraction is performed in each branch network share， and different convolution kernel sizes are used in the deep network part. Separate loss functions are used， and multiple scale gradients are used to adjust the weights of the middle and shallow features simultaneously. Experiments show that the convergence and generalization of each branch network have been improved compared to the single-net network of resnet. At the same time， the algorithm in this paper shows better performance in multi-category data sets， and has practical significance in the big data environment. |
| |
| Keywords: | residual network multi-branch network generalization performance |
|
| 点击此处可从《教育技术导刊》浏览原始摘要信息 |
| 点击此处可从《教育技术导刊》下载免费的PDF全文 |
