基于改进卷积神经网络的机动车图像分类算法  

VEHICLE IMAGE CLASSIFICATION ALGORITHM BASED ON IMPROVED CONVOLUTION NEURAL NETWORK

作　　者：王茜 陈一民[1] 丁友东[1] Wang Qian 1,2,Chen Yimin1,Ding Youdong 1(1College of Computer Engineering and Science, Shanghai University, Shanghai 200072, China;2Information Center, Criminal Investigation Department of Shanghai Public Security Bureau, Shanghai 200083, China)

机构地区：[1]上海大学计算机工程与科学学院,上海200072 [2]上海市公安局刑事侦查总队科技信息科,上海200083

出　　处：《计算机应用与软件》2018年第7期263-266,298,共5页Computer Applications and Software

摘　　要：针对大型数据库的精细化车型分类应用较少、预处理复杂,且识别率不高等情况,提出基于改进卷积神经网络的机动车图像分类算法。算法构建了较之Googlenet V3层级更为简单的神经网络模型;基于该CNN网络,增加了基于样本质心距离的正样本保留方案,在缓解样本不均衡的同时,通过巩固类内边界增强了数据可分性;在网络的全连接层采用了基于神经元重要性分值的dropout方法,在去除无效神经元的同时,提升网络的识别效果。实验结果表明,该算法能更为有效地提取图像特征,较之Googlenet V3算法收敛快,训练耗时短,识别率更高,解决实际问题的能力更强。Aiming at the problems of previous vehicle type classification applications on large-scaled database which are limited in practical cases,complex in pre-processing steps or unsatisfactory in accuracy,this paper proposed image classification algorithm based on improved convolution neural network. The algorithm constructed a simpler CNN network compared with the famous Googlenet V3. While alleviating the problem of sample imbalance,data separability was enhanced by consolidating the inner class boundaries. We applied a dropout method on a full-connecting layer of proposed CNN to remove the useless neurons,and enhance the accuracy of the network. The experimental results show that the algorithm can effectively extract the features of samples. Compared with the Googlenet V3,it gains faster convergence rate,shorter training time and higher accuracy. This method is more capable of solving practical problems.

关 键 词：机动车车型 图像分类 卷积神经网络 难负样本挖掘 DROPOUT 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]