基于特征图切分的轻量级卷积神经网络

卷积神经网络模型所需的存储容量和计算资源远超出移动和嵌入式设备的承载量,因此文中提出轻量级卷积神经网络架构(SFNet).SFNet架构引入切分模块的概念,通过将网络的输出特征图进行"切分"处理,每个特征图片段分别输送给不同大小的卷积核进行卷积运算,将运算得到的特征图拼接后由大小为1×1的卷积核进行通道融合.实验表明,相比目前通用的轻量级卷积神经网络,在卷积核数目及输入特征图通道数相同时,SFNet的参数和计算量更少,分类正确率更高.相比标准卷积,在网络复杂度大幅降低的情况下,切分模块的分类正确率持平甚至更高.

A Lightweight Convolutional Neural Network Architecture with Slice Feature Map

The capacities of mobile and embedded devices are quite inadequate for the requirement of the storage capacity and computational resources of convolutional neural network models. Therefore, a lightweight convolutional neural network architecture, network with slice feature map, named SFNet, is proposed. The concept of slice block is introduced. By performing the “slice” processing on the output feature map of the network, each feature map segment is respectively sent to a convolution kernel of different sizes for convolution operation, and then the obtained feature map is concatenated. A simple 1×1 convolution is utilized to fuse the channels of the feature map. The experiments show that compared with the state-of-the-art lightweight convolutional neural networks, SFNet has fewer parameters and floating-point operations, and higher classification accuracy with the same number of convolution kernels and input feature map channels. Compared with the standard convolution, in the case of a significant reduction in network complexity, the classification accuracy is same or higher.