复杂背景的夜光遥感建筑区检测  

Detection of building area with complex background by night light remote sensing

作　　者：李海[1] 李洋 左峥嵘[1] LI Hai;LI Yang;ZUO Zheng-Rong(National Key Laboratory of Multi-spectral Information Processing Technology,Huazhong University of Science and Technology,Wuhan 430074,China;Institute of Robotics,Shanghai Jiaotong University,Shanghai 200240,China)

机构地区：[1]华中科技大学人工智能与自动化学院多谱信息处理技术国家级重点实验室,湖北武汉430074 [2]上海交通大学机器人研究所,上海200240

出　　处：《红外与毫米波学报》2021年第3期369-380,共12页Journal of Infrared and Millimeter Waves

基　　金：国家自然科学基金(61773389)。

摘　　要：提出了一种新的解决夜光遥感复杂背景问题的单阶段深度卷积检测网络,首先通过提取高维特征再特征选择的思想设计分类网络提取语义特征,并研究不同的通道数网络对降噪的影响;提出灰度能量的先验框匹配,将低噪声高质量的匹配框输入SSD检测网络,并使用积分图思想简化计算;使用可变形卷积以适应目标的形变,并获取更强的几何特征表达能力;通过加入顺序连接与密集连接改进全局语义模块,引入了网络的跨层信息交互,其注意力图综合考虑了高低感受野以有效区分小型目标和背景噪声。在夜光遥感数据集上通过实验验证了所设计的网络相比于其他单阶段网络具有优势,对于复杂背景下的建筑区具有较好的检测效果。A new single-stage deep convolution detection network is proposed to solve the complex background problem of night light remote sensing.Firstly,a classification network is designed by extracting high-dimensional features and then selecting features,and the influence of different channel number networks of noise reduction is studied.A prior box matching of gray-scale energy is proposed,inputting a low-noise and high-quality matching box into SSD detection network,and the idea of integral diagram is used to simplify the calculation.By adding sequential connection and dense connection to improve the global semantic module,the cross layered information interaction of the network is introduced,and its attention map comprehensively considers the high and low receptive fields to effectively distinguish small targets and background noise.Experimental results of the night light remote sensing data set show that the designed network has advantages over the rest single-stage network,which has a better detection effect of the building area under the complex background.

关 键 词：模式识别与智能系统 夜光遥感 深度卷积网络 建筑区检测 复杂背景 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]