《计算机应用研究》|Application Research of Computers

动作识别中基于深度神经网络和GA合并算法的分类决策方法

Classification decision method based on depth neural network and GA merging algorithm in motion recognition

免费全文下载 (已被下载 次)  
获取PDF全文
作者 赵雪章,席运江,黄雄波
机构 1.佛山职业技术学院,广东 佛山 528137;2.华南理工大学,广州 510641
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2019)07-066-2232-05
DOI 10.19734/j.issn.1001-3695.2018.01.0120
摘要 针对人体动作识别中传统方法在分类决策方面存在的问题和缺陷,提出了一种新颖的基于深度神经网络(DNN)和遗传算法(GA)合并算法的非线性分类决策方法。首先提出的合并算法在整个训练集合上对特征提取器进行组合,进而组合成不同的两个独立网络,再利用DNN对两个独立网络进行初始化,进一步利用GA对两个网络进行合并;然后将网络的偏差和权重表示为每层网络间的一个矩阵;最后利用DNN对网络的偏差和权重进行训练,并在合并过程中将矩阵中的每一行当做一个染色体。实验采用了标准MNIST数据集对提出算法的性能进行评估。评估结果显示实验过程中的交叉和突变操作增加了神经元节点,提高了识别性能,并且弱化了不相关和相关神经元节点。因此,提出算法的错误率更低,网络性能更优异。
关键词 动作识别; 分类决策; 重新训练; 遗传算法; 深度神经网络
基金项目 国家自然科学基金面上资助项目(71371077)
佛山市科技计划资助项目(2015AB004241)
本文URL http://www.arocmag.com/article/01-2019-07-066.html
英文标题 Classification decision method based on depth neural network and GA merging algorithm in motion recognition
作者英文名 Zhao Xuezhang, Xi Yunjiang, Huang Xiongbo
机构英文名 1.Foshan Polytechnic,Foshan Guangdong 528137,China;2.South China University of Technology,Guangzhou 510641,China
英文摘要 Aiming at the problems and shortcomings of traditional methods in human motion recognition in classification decision, this paper proposed a novel nonlinear classification decision method based on DNN and GA merge algorithm. Firstly, the proposed merging algorithm combined the feature extractors over the entire training set and combined them into two different independent networks. Then it used DNN to initialize two independent networks and further used GA to merge the two networks. Then it expressed the deviation and weight of the network as a matrix between each layer of the network. Finally, this method used DNN to train the bias and weight of the network, and treated each row in the matrix as a chromosome during the merge process. The experiment used the standard MNIST data set to evaluate the performance of the proposed algorithm. The evaluation results show that the crossover and mutation operations during the experiment increase the neuron nodes, improve the re-cognition performance, and weaken the irrelevant and related neuronal nodes. Therefore, the proposed algorithm has a lower error rate and better network performance.
英文关键词 motion recognition; classification decision; retraining; genetic algorithm(GA); depth neural network(DNN)
参考文献 查看稿件参考文献
 
收稿日期 2018/1/29
修回日期 2018/3/14
页码 2232-2236
中图分类号 TP391.4
文献标志码 A