Abstract: In order to better apply the convolutional neural network model to mobile and embedded devices, it is necessary to reduce the amount of model parameters and reduce computational complexity. Firstly, several popular solutions are briefly introduced. Next, six lightweight convolutional neural network models are elaborated, showing the computational complexity and parameter quantities of different network computing methods. The core building blocks of the model, the overall network structure and innovations are discussed. The classification accuracy of each network and conventional convolutional network on the ImageNet dataset is analyzed. Furthermore, comparing the techniques of lightening the weight of each network, the conclusion is drawn that the direct index is used instead of the indirect index when designing the model. At the same time, the importance of the residual structure to ensure the accuracy of the lightweight model is found. Finally, the development prospect of lightweight convolutional neural network is prospected.

Key words: Convolutional Neural Network（CNN）, lightweight, convolution method

摘要： 为了使卷积神经网络模型更好地应用于移动端和嵌入式设备，必须从减少模型参数量和降低计算复杂度两方面入手。首先简要介绍了目前几种流行的解决方法，并详细阐述了六个轻量化卷积神经网络模型，展示了其中应用的不同网络计算方式的计算量和参数量，论述了模型的核心构建模块、整体网络结构和创新之处。分析了各网络以及常规卷积网络在ImageNet数据集上的分类准确度，进而对比各网络实现轻量化的技巧，得出在进行模型设计时采用直接指标替代间接指标的结论。同时发现了残差结构对保证轻量化模型准确率的重要性。最后对轻量化卷积神经网络的发展前景进行了展望。

关键词: 卷积神经网络（CNN）, 轻量化, 卷积方式