深度卷积神经网络在图像分类任务中取得了极大的成功。现有的基于简化卷积的卷积神经网络结构能够减少网络参数,但会丢失部分特征信息,降低网络性能。为了提高图像分类正确率,提出一种二分支卷积单元。该卷积单元包含两种类型的滤波器,分别用于提取包含特征通道内与通道间信息的特征。以此卷积单元代替传统的滤波器,构建深度卷积神经网络,称为CTsNet。将该网络应用于图像分类任务,在CIFAR10、CIFAR100数据库上进行验证实验。结果表明,二分支卷积单元能够有效提取包含不同信息的特征,增加特征的多样性,减少信息损失,基于二分支卷积单元的CTsNet结构能有效提升图像分类性能。

Deep convolutional neural networks are widely used in the image classification. Current convolutional neural networks architectures based on the simplified convolution can reduce the number of network parameters, but it will lose some of the important information, which decreases the performance of the networks. The two-stream convolutional unit is proposed, in order to improve the accuracy of image classification. The two-stream convolutional unit contains two convolutional filters, which extracts the features containing the information in and across the channels, respectively. Based on the proposed two-stream convolutional unit, a deep convolutional neural network called CTsNet is constructed. Experiments of image classification are conducted on the databases of CIFAR10 and CIFAR100. The experimental results demonstrate that the proposed two-stream convolutional unit can extract features containing the information in and across the channels separately, increase the diversity in features and reduce the information loss. The CTsNet based on the two-stream convolutional unit can improve the recognition performance effectively.