摘要
随着“智慧城市”概念的提出,遥感目标检测逐渐成为城镇的规划、建设和维护的重要方式。为了表征不同城市的差异化遥感特征,解决模型在不同尺度物体上泛化能力不均的问题,提出了一种基于混合分离卷积的金字塔架构搜索方法。首先,分析了遥感图像数据集的空间分布特征;然后,针对其特点构建多感受野混合卷积的搜索空间,进而训练其子网络内的权重;同时借助强化学习算法针对收敛的损失值序列,循环搜索特征提取单元的数量和结构;最终,当架构奖励函数稳定时,固定相应的架构参数和权重矩阵,从而在测试数据上可以自适应融合图像的跨尺度信息,提高同类目标在不同分辨率下的定位精度。该方法搜索出的网络在DIOR遥感数据集上的平均准确率为78.6%,比CornerNet高6百分点,比Cascade R-CNN高1.6百分点,其中小物体准确率比Cascade R-CNN高2.1百分点,并证实了多尺度架构搜索在遥感目标检测的优化能力。
With the concept of“smart city”,remote sensing target detection has gradually become an important way for town planning,construction and maintenance.In order to characterize the differentiating remote sensing features of different cities and solve the problem of uneven generalization ability in the model on objects of different scales,this paper proposes a pyramid structure search method based on hybrid separation and convolution.Firstly,this paper analyzes the spatial distribution characteristics of the remote sensing image dataset,also constructs a multi-receptive field hybrid convolution search space based on its characteristics,and then trains the weights in its sub-network.Secondly,the number and structure of feature extraction units are searched cyclically with the help of reinforcement learning algorithms for the convergent loss value sequence.Finally,when the architecture reward function is stable,the corresponding architecture parameters and weight matrix are fixed,so that the cross-scale information of the image can be adaptively fused on the test data to improve the positioning accuracy of similar targets at different resolutions.The average accuracy of the network searched by this method on the DIOR remote sensing dataset is 78.6%,which is 6 percentage points higher than that of CornerNet,1.6 percentage points higher than that of Cascade R-CNN,and the accuracy of small objects is 2.1 percentage points higher than that of Cascade R-CNN.The optimization ability of multi-scale architecture search in remote sensing target detection was confirmed.
作者
裴婵
廖铁军
PEI Chan;LIAO Tiejun(College of Resources and Environment, Southwest University, Chongqing 400715, China)
出处
《国土资源遥感》
CSCD
北大核心
2020年第4期53-60,共8页
Remote Sensing for Land & Resources
基金
教育部人文规划基金项目“三峡库区农业适度规模经管研究”(编号:15XJA790002)
国家自然科学基金重点项目“土壤中的电场——量子涨落耦合作用”(编号:41530855)共同资助。
关键词
深度学习
遥感检测
架构搜索
图像金字塔
召回率
deep learning
remote sensing detection
architectural search
image pyramid
recall
作者简介
第一作者:裴婵(1996-),女,硕士研究生,主要研究方向为土地资源信息管理,Email:1371566711@qq.com;通信作者:廖铁军(1961-),男,教授,硕士生导师,主要研究方向为土地资源信息管理,Email:ltjhy-007@163.com。
