基于高分辨率可见光遥感图像的建筑物提取技术研究

发布时间：2018-01-03 06:19

本文关键词：基于高分辨率可见光遥感图像的建筑物提取技术研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于大规模可见光遥感图像的全自动建筑物提取技术已经被研究了数十年了。该技术在遥感图像分析领域扮演了重要的角色,它被广泛应用于数字城市,军事侦察,灾害评估等。但是,真实城市场景中建筑物的尺寸范围很宽、颜色纹理十分复杂、日照原因产生的阴影遮挡等难题为屋顶提取带来了巨大地挑战。在本文中,我们提出了一个完整的建筑物区域识别和轮廓估计系统,相比之前的方法,本文的系统既显著地提高了识别准确率,也降低了轮廓估计误差。对于高分辨率遥感图像,整个全自动识别和轮廓提取的过程耗时仅为数秒,成为了当前最快的建筑物提取系统。具体地,本文的贡献可以归纳为如下两个方面:(1)本文提出了一个端到端的多层级融合全卷积神经网络结构,我们称为HF-FCN,该网络结构能够有效地将具有一组差序感知野的神经元感知到的深层图像特征融合起来。该网络结构可以将任意大小的图像不加裁剪或变形地作为网络输入,直接得到最终预测结果。这不仅方便了用户的使用,而且明显减少网络处理时间。经过本文对网络结构的改进,我们的网络参数被压缩至50MB左右。本文还提出了三种HF-FCN的结构变体,并在公开数据集上进行了性能比较实验,证明了这些神经网络结构的合理性。为了验证本文所提网络结构的性能,我们分别和国际上主流的语义分割方法,基于深度学习方法的建筑物识别方法在公开遥感图像数据集上进行了比较,实验结果证明了我们的方案在算法复杂度和识别准确率上具有明显优势。(2)本文设计了一个高效快速的轮廓估计方案,该方案为场景中不同尺寸的建筑物定制了不同的后处理和轮廓估计方案。尤其对于尺寸较大的建筑物区域,本文提出了一种新型的迭代式结构点检测算法,在保证高压缩比的同时,有效地保持住了建筑物的结构特征。本文设计了一组实验充分展示了该算法各环节的耗时情况和轮廓估计结果。为了证明该算法的保结构性,本文以总均方误差和可视化结果为衡量标准,在一组测试集上和多边形逼近算法做了详细的性能比较。实验证明本文的方法更适合处理建筑物轮廓估计任务。
[Abstract]:The extraction technology of automatic building large-scale optical remote sensing image has been studied for decades. Based on this technology in the field of remote sensing image analysis plays an important role, it is widely used in Digital City, military reconnaissance, disaster assessment and so on. However, the size range of building real scene in the city is very wide, color and texture the sunshine is very complex, causes the shadow occlusion problem has brought great challenges for the extraction of the roof. In this paper, we propose a complete building area recognition and contour estimation system, compared to the previous method, this system can significantly improve the recognition accuracy, but also reduces the estimation error for the contour. High resolution remote sensing image, the whole process of automatic recognition and contour extraction time is only a few seconds, become the fastest building extraction system. Specifically, this paper's contribution Can be divided into two aspects as follows: (1) this paper presents a complete end to fusion convolutional neural network structure of multi terminal, we called HF-FCN, the network structure can effectively will have a group of poor deep image feature perception in order to perceive wild neuron fusion to the network structure can be. The image of any size without cutting or deformation as the input of the network, directly to obtain the final prediction result. It is not only convenient for users, but also significantly reduce the network processing time. After the improvement of the network structure, network parameters we are compressed to about 50MB. This paper also puts forward the structure of three HF-FCN variants, and the performance comparison experiment on open dataset, prove the rationality of the structure of neural network. In order to verify the performance of the network structure, we respectively and semantic international mainstream The segmentation method, building recognition method based on deep learning method in the open remote sensing image data sets were compared. The experimental results show that our scheme has obvious advantages in the complexity of the algorithm and the recognition accuracy rate. (2) this paper designs an efficient contour estimation scheme, the scheme for different sizes the building scene customized postprocessing and different contour estimation scheme. Especially for building large size region, this paper presents a new iterative structure detection algorithm, in order to ensure the high compression ratio at the same time, effectively maintain the structural features of the building. This paper designs a set of experiments demonstrating the time and the outline of the link of the estimation results. In order to protect the structure of the algorithm is proved, based on the total mean square error and the visualization results as the standard, in a set of test set The performance comparison between the upper and the polygon approximation algorithms is made in detail. The experiments show that the proposed method is more suitable for the building profile estimation task.

【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】