用于大规模深度卷积分类网络DOA估计的标签分解方法  

Deep Convolution Neural Network Label Decomposition Method for Large Scale DOA Estimation

作　　者：吴双 袁野 马育红 黄敬健 袁乃昌[1] WU Shuang;YUAN Ye;MA Yuhong;HUANG Jingjian;YUAN Naichang(The State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System,National University of Defense Technology,Changsha,Hunan 410073,China)

机构地区：[1]国防科技大学电子信息系统复杂电磁环境效应国家重点实验室,湖南长沙410073

出　　处：《信号处理》2021年第1期1-10,共10页Journal of Signal Processing

摘　　要：为了有效地解决使用深度神经网络求解波达方向(DOA)估计涉及到的大规模分类器的训练和部署实现,本文提出将传统的one-hot分类器分解为多个类别互质的小分类器,然后联合使用多个互质分类器的分类结果重构原始one-hot标签。首先使用标签分解,将原始标签分解为多个互质的小标签,小标签对应的类别为原始类别对质数取余数的结果。其次,通过独立并行地训练每一个互质分类器,降低了大类别条件下分类器的训练难度。仿真结果表明,相比one-hot分类器,互质分类器网络的复杂度低,易于训练。另外,使用互质分类器进行DOA估计能够实现超分辨并且估计的精度比one-hot分类器以及稀疏贝叶斯学习等方法更高。In order to effectively solve the training and deployment implementation of large-scale classifiers involved in deep neural network to solve direction of arrival(DOA)estimation,we propose to decompose the traditional one-hot classifier into small co-prime classifiers with relatively small label,and then reconstruct the original one-hot label by combining the classification results of co-prime classifiers.First,the original label is decomposed into a number of small co-prime labels.The corresponding category of the small labels is the result of the remainder of the prime number of the original label.Secondly,by training each co-prime classifier independently and in parallel,the difficulty of training classifiers under the condition of large categories is reduced.The simulation results show that compared with the one-hot classifier,the co-prime classifiers has lower complexity and is easy to train.In addition,DOA estimation using co-prime classifiers can achieve super-resolution and the estimation accuracy is higher than that of one-hot classifier and sparse Bayesian learning method.

关 键 词：波达方向估计 深度卷积神经网络 标签分解 大规模分类问题 

分 类 号：TN911.7[电子电信—通信与信息系统]