一种神经网络架构搜索遥感图像分类方法  

A Network Architecture Search Remote Sensing Image Classification Method

作　　者：景维鹏[1] 张明伟 林敬博 JING Wei-peng;ZHANG Ming-wei;LIN Jing-bo(College of Information and Computer Engineering, Northeast Forestry University, Harbin 150040,China)

机构地区：[1]东北林业大学信息与计算机工程学院,哈尔滨150040

出　　处：《哈尔滨理工大学学报》2021年第1期25-29,共5页Journal of Harbin University of Science and Technology

基　　金：教育部大学生创新训练计划项目(201910225161);国家自然科学基金(31770768);2018-2019年黑龙江省档案局档案科技项目计划(HDK2018-20).

摘　　要：针对传统深度卷积神经网络遥感图像分类方法需要不断进行试错,花费相关专家大量时间和计算资源的问题,提出了一种基于神经网络架构搜索的遥感图像分类方法。首先搜索最优单元(cell),再以预定义的方式堆叠该最优单元得到目标网络。其中该方法将离散的网络架构表达连续化,从而可以使用梯度下降的方法在离散的搜索空间内进行搜索,实现其高效性。为了提高其精确性,该方法还对目标网络在训练集上进行重新训练。实验结果表明,在实验测试集上测试用时4 h训练所得模型分类准确率达到了88.57%,说明了该方法的高效性以及高精确性。The traditional deep convolutional neural network remote sensing image classification method requires trial and error many times,requiring a lot of time and computing resources.This paper proposes a remote sensing image classification method based on neural network architecture search.The method first searches for an optimal cell,and then stacks the optimal cell in a predefined manner to obtain a target network.The method continues relaxation of the architecture represenation,so that the gradient descent method can be used to search in the discrete search space to achieve high efficiency.In order to improve its accuracy,we also retrain the target network on the training set.The experimental results show that the classification accuracy of the model obtained by the test in the experimental test set is 88.57%,showing the high efficiency and high accuracy of the method.

关 键 词：深度学习 高分辨率遥感图像 卷积神经网络 网络架构搜索 图像分类 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]