多尺度残差网络模型的研究及其应用  

Research and application of the multi-scale residual networks model

作　　者：王飞[1] 张莹[1,2] 卲豪 张东波 牟清萍[1] Wang Fei;Zhang Ying;Shao Hao;Zhang Dongbo;Mu Qingping(College of Information Engineering,Xiangtan University,Xiangtan 411105,China;National Engineering Laboratory for Robot Visual Perception and Control Technology,Changsha 410082,China)

机构地区：[1]湘潭大学信息工程学院,湘潭411105 [2]机器人视觉感知与控制技术国家工程实验室,长沙410082

出　　处：《电子测量与仪器学报》2019年第4期19-28,共10页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(61773330);湖南省重点学科基金;湖南省自然科学基金(2017JJ2251)资助项目

摘　　要：针对传统的卷积神经网络不能充分利用图像的多尺度信息,以及网络层数的增加导致优化参数增加的问题。提出多尺度残差网络模型通过多尺度跨通道的卷积融合提高图像的特征表征能力;然后把大的卷积核分解成小的非对称卷积核降低网络的参数计算;接着利用残差网络原理来降低深层网络的梯度消失问题;最后将提出的多尺度卷积模块嵌入到Lenet网络中。在Mnist数据集上的测试结果证明分类准确率比原始Lenet网络提高了0. 24%,在LFW数据集上的测试结果表明分类准确率优于Deep Face、Web Face等传统算法。For the traditional convolutional neural network,the optimization parameter will increase as the number of network layer increases. Also,the multi-scale information of the image cannot be made full of use. To solve these problems,the multi-scale residual networks model is proposed to improve the feature characterization through convolution fusion of multiscale and cross channel. Then the large convolution kernel is decomposed into a small asymmetric convolution kernel in order to reduce network parameter calculation,the residual networks is used to reduce the impact of the gradient disappearance. Finally,The proposed multi-scale convolutional module is embedded in Lenet networkfor image classification. The experimental results on the Mnist dataset show that the classification accuracy is improved 0. 24% than the original Lenet network.The experimental results on the LFW dataset show that the classification accuracy of the proposed model is better than Deep Face,Web Face and other traditional algorithms.

关 键 词：多尺度残差网络 卷积神经网络 跨通道卷积 核分解 

分 类 号：TP391.73[自动化与计算机技术—计算机应用技术] TN911.41[自动化与计算机技术—计算机科学与技术]