【作者】 刘博；

【导师】 张钧萍；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2023， 硕士

【摘要】 无人机多光谱遥感在农业中有着广泛的应用,例如水稻长势监测、产量预估和大面积性状检测等。无人机获取的多光谱图像具有较高的空间和时间分辨率,但是多光谱相机拍摄的单幅图片所包含的信息量较少,如何将这些图像高精度、快速拼接成一幅范围更大、包含更多信息量的图像具有重要的研究意义。本论文对无人机多光谱遥感图像配准与拼接方法的相关理论进行了的研究与分析,围绕着特征提取、线段检测与校正、快速拼接等关键问题展开了研究,获得了阶段性的成果,便于后续大规模场景下农事活动的处理与分析。针对现有的匹配算法在3D视角变化时匹配对数量少、配准精度低等问题,本论文研究使用深度学习网络进行无人机多光谱图像配准。首先采用SuperPoint网络进行特征提取,并行得到特征点的位置信息和特征描述符,然后馈送到SuperGlue网络进行特征匹配,计算变换矩阵后,经过图像变换得到配准结果。实验结果表明采用的方法具有更高的检测效率和匹配性能,得到了更为准确的配准结果。针对全色图像或RGB图像拼接有着全面广泛的研究,然而对于多光谱图像的精确拼接方法较少;如何减少拼接结果产生失真和变形也是研究的难点之一。因此提出一种有效的无人机多光谱图像拼接方法,首先应用第一部分的深度学习配准框架进行图像之间的配准,再利用APAP算法进行图像拼接,并基于线段检测器和多点共线约束进行线段检测和校正,这可以保护农田地区独特的线段结构,获得更为准确的拼接结果。实验结果表明,所提出的方法获得了优于其他典型方法的拼接性能,拼接结果的精度更高。在农业生产的一些特定场景中,例如水稻扬花期的性状检测,这个时期的时间窗口比较窄,对于拼接的实时性有更高的要求。因此提出一种有效的多幅无人机多光谱图像快速拼接方法。首先计算相邻输入的多光谱遥感图像的重叠度,利用增量搜索算法筛选出最佳图像拼接路径并记录保存,采用基于接缝驱动的拼接方法来消除重影,并利用渐入渐出算法进行图像融合,得到多光谱遥感图像的拼接结果。实验结果表明,对比常用的图像拼接软件和经典拼接算法等,论文采用的方法在速度上达到了比较先进的性能,优于其他方法。更多还原

【Abstract】 UAV multispectral remote sensing has a wide range of applications in agriculture,such as rice growth monitoring,yield prediction and large area trait detection.The multispectral images acquired by UAVs have high spatial and temporal resolution,but the single images taken by multispectral cameras contain less information.It is important to study how to stitch these images into an image with greater range and more information with high accuracy and speed.We have conducted a thorough analysis of theories concerning UAV multispectral remote sensing image registration and stitching,concentrating on key topics such as feature extraction,line segment detection and correction,and rapid stitching.The results of this study have been used to aid in the processing and analysis of agricultural activities in the following large-scale scenes.In response to the existing matching algorithms with low number of matching pairs and low registration accuracy when the 3D viewpoint changes,this thesis uses deep learning networks for UAV multispectral image registration.Firstly,the SuperPoint network is used for feature extraction,and the location information and feature descriptors of feature points are obtained in parallel,and then fed to the SuperGlue network for feature matching,and after calculating the transformation matrix,the registration results are obtained after image transformation.Experimental findings demonstrate that the adopted technique has a more potent detection efficiency and matching performance,as well as a more precise registration outcome.There are comprehensive and extensive research results on the stitching of panchromatic or RGB images,but there are few methods for accurate stitching of multispectral images;how to reduce the distortion and deformation of stitching results is also one of the difficulties in the research.Therefore,an effective UAV multispectral image stitching method is proposed,which firstly applies the deep learning alignment framework in the first part to align between images,then uses the APAP algorithm for image stitching,and performs line segment detection and correction based on line segment detector and multi-point co-linear constraint,which can protect the unique line segment structure of farmland areas and obtain more accurate stitching results.The experimental results show that the proposed method obtains better stitching performance than other typical methods and higher accuracy of stitching results.In some specific scenarios of agricultural production,such as trait detection during the flowering period of rice,the time window of this period is narrower,and there is a higher requirement for the real-time performance of stitching.Therefore,an effective fast stitching method for multiple UAV multispectral images is proposed.Firstly,the overlap degree of adjacent input multispectral remote sensing images is calculated,the best image stitching path is filtered out using incremental search algorithm and recorded and saved,the stitching method based on seam driven is used to eliminate ghosting,and the image fusion is carried out using fading-in and fading-out algorithm to obtain the stitching results of multispectral remote sensing images.Experimental results demonstrate that the technique employed in the paper surpasses other techniques in terms of speed and surpasses commonly employed image stitching software and classical stitching algorithms,etc.更多还原