高分辨率遥感影像复杂场景下密集建筑物提取研究

Robust and automatic building detection and contour extraction in high speed under complex region is one of the key problems in remote sensing image processing and interpretation. Which can be used widely in applications such as map updating, disaster reduction and disaster relief, military and national defense. Traditional methods have poor levels of automation and cannot adaptive to various sensors. More importantly, most of them can hardly be applied to real, complex remote sensing imageries, which makes a large volume of remote sensing data unused. Based on previous studies, in this proposal we propose to develop new methods for building extraction for complex and density building regions. Firstly, a top-down saliency model is developed for built-up area extraction. In which, we model human visual selective attention mechanism, and to better segment built-up area from backgrounds human level built-up priors are added to the model. Then, a new method combining object-oriented methodology with deep convolutional neural networks is proposed to detection building in the built-up areas. In which, the usage of object-oriented methodology is to generate more valid object proposals than conventional proposal methods. Using this new proposal method, the proposed method can detective building more accurately. Finally, we use a new method called Bayesian Program Learning (BPL) to generate geometry contour of buildings. BPL can learn human level concept from a few examples. After building geometry contour models, BPL is applied to obtain building’s geometry contour. This research may help to develop new methods for remote sensing image intelligent interpretation, and give positive effective on image processing, computer vision.

复杂场景下快速、鲁棒的建筑物自动检测及轮廓提取是当前遥感影像智能解译的关键问题之一，在地图更新、减灾救灾、军事国防等方面具有广阔的应用前景。传统方法的自动化程度较差，对不同传感器影像的自适应性较弱，且在复杂场景下的鲁棒性较低，限制了遥感影像的有效利用。本项目拟在前期遥感影像目标检测工作的基础上，针对复杂场景下密集建筑提取进行研究。首先，以遥感影像建成区检测为导向，借鉴人的视觉选择性注意机制，构建自顶向下的视觉显著性模型，实现建成区的有效检测；在此基础上，将面向对象思想与深度卷积神经网络有机结合，通过有效的似物性窗口生成方法，提高密集区域单体建筑物的检测精度；最后，模仿人脑高层分析功能，建立屋顶轮廓的二维几何模型，利用贝叶斯概念学习方法，实现建筑物的轮廓提取。相关研究成果有望推动遥感影像智能解译的发展，并对图像处理、计算机视觉等相关领域产生积极的影响。

建筑物是遥感图像里一类非常重要的目标。建筑物检测在地图更新、减灾救灾、军事国防等方面具有广阔的应用前景。传统方法的自动化程度较差，对不同传感器影像的自适应性较弱，且在复杂场景下的鲁棒性较低，限制了遥感影像的有效利用。针对这些难题，本项目开展了基于深度学习技术的遥感图像建筑物及地面目标检测技术研究，标注并发布了迄今为止最大的基于国产高分2号卫星的大规模建筑物检测数据集5M-Building，研究并实现了基于多尺度特征嵌入的目标检测模型、研究了遥感影像全色和多光谱融合算法、并引入生成对抗模型设计了面向建筑物变化检测的网络模型，等等一系列研究成果，发表在IEEE Transactions on Geoscience and Remote Sensing, IEEE Geoscience and Remote Sensing Letters，ICASSP等遥感领域知名期刊和学术会议上。相关研究成果受到领域内同行的认可，Google Scholar 引用400余次，并产生高被引论文一篇。与此同时，相关研究成果已经成功应用到应急管理部减灾中心业务中，提升了业务效率。