Abstract: Aiming at the problems of more and more complex structure of deep convolutional neural network and the large parameters, a new structure of compact convolutional neural network Width-MixedNet and its multi-branch basic module Conv-mixed are designed. The architecture extends the width of the convolutional neural network. The branching structure of Conv-mixed makes multiple different convolutional layers deal with the same feature map and extract different features. In the recognition task of the deep convolutional neural network, the superposition of multiple small convolutional layers is used to reduce the feature maps layer by layer instead of the full connection layers, which is for final feature extraction. The number of parameters of Width-MixedNet is only 3.4×105, which is only 1/30 of the traditional deep convolutional neural networks. Experiments are conducted at CIFAR-10, CIFAR-100 and MNIST with accuracy rates of 93.02%, 66.19% and 99.59%. The experiments show that Width-MixedNet has better performance and learning ability, which observably reduces the parameter size of the network and improves the recognition accuracy.

Key words: deep learning, convolutional neural network, compact structure, width expansion, image recognition

摘要： 针对深度卷积神经网络的结构越来越复杂，参数规模过于庞大的问题，设计出一种新的紧凑型卷积神经网络结构Width-MixedNet和其多分支的基本模块Conv-mixed，该架构扩展了卷积神经网络的宽度。Conv-mixed利用分支结构使多个不同的卷积层处理同一个特征图，提取不同的特征。在深度卷积神经网络的识别任务中，使用多个小型卷积层叠加，逐层缩小特征图的方法代替全连接层进行最后的特征提取。整个Width-MixedNet架构的参数数量只有3.4×105，仅有传统深度卷积神经网络的1/30。分别在CIFAR-10、CIFAR-100和MNIST数据集上进行实验，准确率分别达到了93.02%、66.19%和99.59%。实验表明，Width-MixedNet有更强的学习能力和表现能力，在提高识别精度的同时，大大降低了网络的参数规模。

关键词: 深度学习, 卷积神经网络, 紧凑型结构, 宽度扩展, 图像识别