【作者】 任才俊；

【导师】 陈欢欢；

【作者基本信息】 中国科学技术大学 ， 计算机应用技术， 2022， 硕士

【摘要】 变化检测作为城市规划管理工作的重要一环,其目标是定位于不同时间在同一区域中发生变化的部分,变化检测是后续分析管理工作的基础。近年来随着互联网技术的飞速发展,信息数据的采集与应用变得更加便捷,同时数据的精度与规模都在不断提升。对于城市规划管理来说,高精度、大规模的文本及图像数据都有助于提高相应任务的结果,但同时也对处理数据的能力与方法提出了更高的要求。遥感图像变化检测是基于遥感设备拍摄的城市图像开展的图像处理工作,任务的主要难点来自于数据规模与数据精度两方面。由于数据规模极大,采用人工标注的方式对已有数据进行处理会消耗很长的时间,导致相应的数据缺少人工加持的精准标注,需要用到无监督算法。同时,考虑到图像数据的分辨率较高,位于同一地点的同一建筑物,由于拍摄条件的不同,在图像中会呈现出一定的差异,形成噪音,对准确定位变化区域造成了干扰。本文在无监督条件下,为了减少噪音的干扰提高检测准确度,分别提出了基于机器学习与深度学习的两种遥感图像变化检测方法,主要工作总结为以下两点:(1)提出了一种基于搜索匹配与降维聚类的变化检测算法。该方法将图像中颜色差距不大的相邻像素点合并为一个超级像素,之后在一组处理过后的图像上依据位置与像素值进行匹配,将匹配成功的像素点认为是未变化的相同区域,利用类搜索算法将相应区域排除,而余下的区域被认为是可能产生变化的区域。最后利用降维方法对余下区域的差值进行处理,使用聚类方法对像素点进行分组,取像素点个数最少的组认为是最终变化区域。该方法满足实际应用中处理要求,在面对规模较大的数据时,能较好地适应实际生产应用需求;(2)鉴于深度神经网络在无监督学习方面的成功应用,本文进一步提出了基于生成对抗网络的变化检测方法。利用训练生成对抗网络对原有的遥感图像进行再生成处理,可以生成任意数量的相应图像。通过对网络结构、目标函数以及训练数据的重新设计,可以使得在生成图片中,主要变化区域依然保持变化,而噪音干扰区域保持一致。因此直接利用生成图片的平均差值结果进行阈值分割处理,就可以得到最后的主要变化区域。该方法在面对复杂的城市场景以及无变化场景时,相较于其他同类型变化检测方法,取得了较为明显的准确度提升。本文提出的算法已被应用于相应的城市规划变化检测任务中,对不同场景下的遥感图像数据的实验结果证明了本文算法的有效性和实用性。更多还原

【Abstract】 As an important part of urban planning and management,change detection aims to locate the changed areas in the same place at different time,and is an significant basis for subsequent analysis and management.With the rapid development of Internet technology,the collection and application of information data has become more convenient,the accuracy and scale of data are constantly improving.For urban planning and management,text and image data with high-precision and large-scale are helpful to improve the efficiency,however,it also puts forward higher requirements on the ability and method of processing data.Change detection in remote sensing image is based on the data obtained by satellites.The main difficulties come from two aspects:data scale and data accuracy.Given that images contain the entire city,the scale of data is extremely large.It required much human intervention to identify changes in so many images.Therefore,the change detection methods are always unsupervised algorithms without labels.Meanwhile,due to the high accuracy of the image data,the same building located in the same place will show some differences in paired images due to the different shooting conditions,which would become noises in change detection.The existence of these noises interferes with the actual changed areas.Under unsupervised conditions,two remote sensing images change detection methods based on machine learning and deep learning are proposed separately to reduce noise interference and improve the accuracy of change detection.(1)A change detection method based on matching and clustering is proposed.This method merges adjacent pixels with small color gaps into one super pixel,and then perform matching procedure on the paired images according to the positions and pixel values,and considers the matched pixels as the centre of unchanged areas.Then a searching-like procedure is applied to exclude the unchanged areas,and the remaining areas are considered to be possible changed areas.Finally,the Principal Component Analysis is performed to process the remaining area,the clustering method is used to group the pixels,and the group with least pixels is considered as the results.This method has an advantage of fast calculation,it could satisfy the requirements of real-world application when dealing with large-scale data.(2)In view of the successful application of neural networks in unsupervised learning,we further propose a change detection method based on Generative Adversarial Networks.The original remote sensing images could be regenerated by the trained network model.Based on the original paired images,any number of corresponding images could be generated.At the same time,by designing the network structure,the objective function and the method of obtaining training data,it is possible to keep the main change area in the generated images,while the noises remain the same.Therefore,directly performing threshold segmentation on the average difference of the generated images could obtain the final results.Comparing to other deep-learning methods.the proposed method achieved obvious improvement on accuracy when processing complex rural scenes and scenes with no changes.Two methods have been applied to relative change detection tasks.Experiments on the images data describing different situations have proved the efficiency and practicability of proposed methods.更多还原