摘要
针对目标识别需求,对基于神经网络的深度学习方法展开研究。由于深度学习模型中包含了对数据的先验假设,因此人工设计神经网络需要领域内专家丰富的先验知识,且具有劳动密集与时间成本高的缺点。为了获得超越专家个人经验、表现更好的网络,采用一种可微神经结构搜索的高效结构搜索方法,将搜索空间放宽为连续的空间,然后通过梯度下降来优化体系结构的验证集性能,从而找到面向目标识别的最优神经网络结构。仿真实验结果表明,将基于神经网络结构搜索的目标识别方法应用于“低慢小”类目标识别是可行的。
In view of the requirements of target recognition,the deep learning methods based on neural network are set off.Generally,there is a priori hypothesis of data contained in the in-depth learning model,the artificial design aimed at neural network for data needs an abundance of priori knowledge for experts in the field,and has the disadvantages of labor-intensive and high time cost.In order to obtain better network performance beyond the personal experience of network design experts,an efficient structure search method,i.e.Differentiable Architecture Search,is adopted.In this method,the search space is broadened to be continuous,and then the performance of the verification set of the architecture is optimized by gradient descent,finding the optimal neural network structure for target recognition.The simulation results show that it is feasible to apply the target recognition method based on neural network structure search to the LSS target recognition.
作者
卞伟伟
邱旭阳
申研
BIAN Weiwei;QIU Xuyang;SHEN Yan(Beijing Institute of Mechanical Equipment,Beijing 100854,China)
出处
《空军工程大学学报（自然科学版）》
CSCD
北大核心
2020年第4期88-92,共5页
Journal of Air Force Engineering University(Natural Science Edition)
基金
国防基础科研计划(JCKY2016204A601)。
关键词
目标识别
卷积神经网络
神经网络结构搜索
深度学习
target recognition
convolution neural network
neural network structure search
deep learning
