电子测量与仪器学报

李帷韬,韩慧慧,焦点,汤健,丁美双.基于深度迁移学习的大雾等级智能认知方法研究[J].电子测量与仪器学报,2020,34(2):88-96

基于深度迁移学习的大雾等级智能认知方法研究

Research on intelligent cognition method of fog level based on deep transfer learning

DOI：

英文关键词:fog grade interleaved group convolutions deep stochastic conguration networks entropy of semantic error transfer learning

基金项目:异构网络和异构数据融合的变电站一体化智能运维关键技术研究与应用资助项目

作者	单位
李帷韬	1.合肥工业大学电气与自动化工程学院
韩慧慧	1.合肥工业大学电气与自动化工程学院
焦点	1.合肥工业大学电气与自动化工程学院
汤健	2.北京工业大学信息学部
丁美双	3.合肥共达职业技术学院

Author	Institution
Li Weitao	1.School of Electric Engineering and Automation, Hefei University of Technology
Han Huihui	1.School of Electric Engineering and Automation, Hefei University of Technology
Jiao Dian	1.School of Electric Engineering and Automation, Hefei University of Technology
Tang Jian	2.Faculty of Information Technology, Beijing University of Technology
Ding Meishuang	3.Hefei Gong Da Vocational and Technical College

摘要点击次数: 491

全文下载次数: 1

中文摘要:

已有雾气检测模型存在缺陷以及深层神经网络优化存在困难。模仿人类认知模式，借鉴迁移学习和闭环控制理论思想，探索了一种基于动态交错组卷积的雾气等级智能认知方法。首先，构建交错组卷积层以降低卷积运算的通道冗余；其次，构建可区分性测度指标和认知决策信息系统，获取雾气图像多层次差异化简约特征空间数据结构；再次，设计深层随机配置网络分类器，构建具有强泛化能力的分类准则；最后，基于广义误差和熵理论, 仿人类反复推敲比对的认知模式实时评测雾气等级认知结果的可信度，基于迁移学习机制实现雾气图像多层次差异化特征空间及其分类准则的自寻优重构，对可信度低的雾气图像进行再认知。15 000幅大雾图像的平均识别率为9598%，实验结果表明，所用方法与其他算法相比，增强了模型的泛化能力，提升了模型的认知精度。

英文摘要:

Existing fog detection models have defects and the optimization of deep neural network is difficult. Based on imitating human cognitive model, transfer learning and closed loop control theory, this paper explores an intelligent cognitive method of fog level based on dynamic interleaved group convolution. Firstly, an interleaved group convolution layer is constructed to reduce the channel redundancy of convolution operation. Secondly, a discriminability measure index and a cognitive decision information system are constructed to obtain the spatial data structure of differentiated and simplified features of fog images. Thirdly, the deep stochastic conguration networks classifier is designed and the classification criterion with strong generalization ability is constructed. Finally, based on the generalized error and entropy theory, the cognitive model that mimes human repeatedly deliberates and compares to evaluate the credibility of the cognitive results of fog grades in real time. Based on the transfer learning mechanism, the self optimization reconstruction of the multi level differentiated feature space of fog images and their classification criteria are realized, and the low credibility fog images are re recognized again. The average recognition rate of 15,000 fog images is 9598%. The experimental results show that compared with other algorithms, this method enhances the generalization ability and improves the cognitive accuracy of the model.

查看全文查看/发表评论下载PDF阅读器