场景解译框架下的高速铁路沿线建筑物隐患自动识别

发布时间：2018-05-04 03:30

本文选题：高铁沿线隐患 + 场景解译　；参考：《西南交通大学》2017年硕士论文

【摘要】：高速铁路沿线地理环境复杂,存在诸多安全隐患,尤其是沿线的房屋、工厂等非法建筑严重影响到高铁的安全运营。因此,必须及时对高铁沿线建筑隐患进行排查。传统的人工实地勘察检测建筑隐患的方法费时费力、效率低下,难以对整个高铁网络进行有效的监控。高分辨率遥感技术具有实时性、周期性等技术优势,为快速、客观、动态地监测高铁沿线建筑隐患提供了有效的技术手段。高分辨率遥感影像地物细节信息十分丰富,但是也存在大量"同物异谱"和"异物同谱"的现象,导致传统的基于像素提取建筑方法精度较低。面向对象方法由于顾及了像素之间的空间关系,在一定程度上提高了高分辨率遥感影像建筑物识别的精度,但是确定最优分割尺度往往比较困难。此外,这两种方法提取建筑时往往基于图像的底层视觉特征分析,并没有建立在图像所描述的高层次语义特征之上,因此,存在明显的语义鸿沟,影响建筑物识别的精度。为了突破这一限制,需要从更高层次的场景层次去理解高分辨率遥感影像。本文选取京沪高铁宿州-蚌埠某段的Google Earth影像作为研究数据。在场景解译框架下,本文首先建立高铁沿线影像块样本库,然后将高铁沿线影像划分为重叠的影像块。将影像块看作文档,通过视觉词袋模型和潜在狄利克雷分布主题模型分别得到影像块的视觉单词直方图表示和潜在的语义主题混合比例信息,输入到SVM分类器得到影像块的类别,最后通过类别投票法确定每个像素的类别,从而实现建筑物隐患的自动识别;将影像块输入到经过训练的卷积神经网络,通过卷积、池化、全连接操作得到全连接层,输入到Softmax得到每个影像块的类别概率分布,最后通过等权平均的方式得到每个像素的类别概率分布,取概率最大所属类别作为该像素的类别,从而实现建筑物隐患的提取。经过实验分析,得到以下结论:经过实验分析,得到以下结论:(1)相比于传统的基于像素和面向对象使用底层特征的建筑隐患识别方法,基于场景解译方法能够显著提升结果紧凑性和完整性,总体精度和生产者精度最高均可达91%,kappa系数可达0.71,与地面真实值较为接近;(2)场景解译框架下各方法中,卷积神经网络方法通过自主学习,避免了视觉词袋模型和主题模型人工设计特征的局限性与盲目性,目视评价与指标评价上表现最优。
[Abstract]:Because of the complex geographical environment along the high-speed railway, there are many hidden dangers in safety, especially the illegal buildings such as houses and factories along the high-speed railway seriously affect the safe operation of high-speed railway. Therefore, must carry on the investigation in time to the construction hidden danger along the high-speed line. The traditional manual investigation and detection method of building hidden trouble is time-consuming and inefficient, so it is difficult to monitor the whole high-speed railway network effectively. High-resolution remote sensing technology has the advantages of real-time and periodicity, which provides an effective technical means for rapid, objective and dynamic monitoring of building hidden trouble along high-speed railway line. High resolution remote sensing images are rich in detailed information of features, but there are also a large number of "isospectral" and "foreign body isospectral" phenomena, which leads to the low accuracy of traditional methods of pixel extraction and construction. Due to the spatial relationship between pixels, the object-oriented method improves the accuracy of building recognition in high-resolution remote sensing images to a certain extent, but it is often difficult to determine the optimal segmentation scale. In addition, these two methods are often based on the analysis of the underlying visual features of the image, and are not based on the high-level semantic features described by the image. Therefore, there is a clear semantic gap, which affects the accuracy of building recognition. In order to overcome this limitation, high resolution remote sensing images need to be understood at a higher level. In this paper, the Google Earth image of Suzhou-Bengbu section of Beijing-Shanghai high-speed train is selected as the research data. Under the frame of scene interpretation, this paper first establishes the sample database of high-speed railway image blocks, and then divides the high-speed railway images into overlapping image blocks. The image block is regarded as a document, and the visual word histogram representation and the potential semantic topic mixed proportion information are obtained by the visual word bag model and the potential Drickley distribution theme model, respectively. Input into the SVM classifier to get the classification of the image block, finally determine the category of each pixel through the class voting method, thus realize the automatic identification of the hidden trouble of the building; input the image block to the trained convolution neural network, through the convolution, pool, pool, The full connection layer is obtained by the full join operation, and the class probability distribution of each image block is obtained by input into Softmax. Finally, the class probability distribution of each pixel is obtained by equal weight average, and the category belonging to the maximum probability is taken as the class of the pixel. In order to achieve the extraction of hidden dangers of buildings. Through the experimental analysis, the following conclusions are obtained: through the experimental analysis, the following conclusions are obtained: compared with the traditional method based on pixel and object-oriented using the bottom feature of building hidden trouble recognition, Based on the method of scene interpretation, the compactness and integrality of the result can be significantly improved. The highest overall precision and producer precision can reach 91kappa coefficient of 0.71, which is close to the real value of ground level. The convolution neural network method avoids the limitation and blindness of the artificial design features of visual word bag model and theme model through autonomous learning. The visual evaluation and index evaluation are the best.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U298;TP751

【参考文献】