非线性激活的聚合残差神经网络汽车胎纹识别  

Tire tread pattern recognition based on non-linear activated aggregation residual neural network

作　　者：陈德海 潘韦驰 马原 黄艳国 CHE Dehai;PAN Weichi;MA Yuan;HUANG Yanguo(School of Electrical Engineering and Automation,Jiangxi Universityof Scienceand Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程及自动化学院

出　　处：《江西理工大学学报》2019年第5期80-85,共6页Journal of Jiangxi University of Science and Technology

基　　金：国家自然科学基金资助项目(61463020);江西省自然科学基金资助项目(20151BAB206034)

摘　　要：针对汽车胎纹人工提取特征手段复杂、识别困难等问题,提出一种采用基数维度变换的聚合残差神经网络进行汽车胎纹识别的方法.在ResNeXt-50网络的基础上,对原有网络结构进行压缩,减少了聚合残差单元的基数以及瓶颈宽度,同时引入非线性激活函数Swish,加强网络模型的收敛能力并提高准确率.使得模型在保持识别汽车胎纹的能力的同时,大幅压缩了参数量并提升了识别汽车胎纹的能力.通过理论分析与实验,验证了方法的有效性.In order to solve the problems such as the complexity of manual feature extraction and the difficulty of identification,a method of identifying automobile tire marks by using the aggregate residual neural network based on the transformation of cardinality dimension is proposed.On the basis of ResNeXt-50 network,the original network structure is compressed to reduce the cardinal number and bottleneck width of aggregate residual units.And the non-linear activation function Swish is introduced to enhance the convergence ability and accuracy of the network model.In this way,the model can greatly reduce the number of parameters and improve the ability of recognizing automobile tire marks while maintaining the ability of recognizing automobile tire marks.Through theoretical analysis and experiments,the effectiveness of the method is verified.

关 键 词：汽车胎纹识别 深度学习 残差神经网络 激活函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]