【作者】 曾峦；

【导师】 谭久彬；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2012， 博士

【摘要】 遥感图像拼接与记录技术是获取高分辨率、大视野的战场图像侦察信息的有效手段，是现代信息化侦察装备研究领域的关键技术之一。但现有图像拼接方法存在稳定性不好、速度慢、自动化程度不高、适应性不强和特征描述符构建复杂、描述向量维数过高等弱点，以及现有数据记录方法持续记录速率不够高的问题，直接影响到空天侦察能力的持续提升。课题“基于不变特征的图像拼接及软同步直写硬盘记录技术研究”的目的是：对高分辨率成像传感器获取的图像实施全自动拼接、实时持续记录等技术进行深入的研究，探讨一种适用于大幅面遥感图像的全自动拼接方法、在恶劣环境下可靠工作的高速长时持续记录系统的设计和实现方法，为我国遥感侦察系统性能的进一步提高奠定理论和技术基础。本课题的研究，在空间探测、航天遥感数据处理、雷达成像数据处理、SAR对抗研究等领域也有广泛的应用前景。本文首先基于光学成像的基本原理，全面总结了图像拼接系统的基本结构，推导了图像变换的数学模型，系统地阐述了图像匹配、图像校正、图像变换和图像融合等关键技术，为展开大幅面图像拼接技术的研究奠定理论基础。在深入研究常规的图像特征提取技术的基础上，引入尺度不变特征变换（Scale Invariant Feature Transform，SIFT）技术，用于大幅面遥感图像的特征提取。针对SIFT描述符的不变性不够完善和描述向量维数过高的问题，提出了一种基于扇形区域分割的SIFT描述符构建方法，取消了构建描述符时对区域的旋转操作，把128维的描述向量降低到48维。针对SIFT描述符对低对比度图像适应能力差的弱点，提出了一种自适应对比度阈值控制提取关键点的方法，增强了SIFT算法对遥感图像的适应能力，扩展了其应用范围。根据SIFT特征多量性和遥感图像幅面大的特点，提出了基于统计、变换模型和改进的RANSAC等三种自动匹配策略，能依据输入图像自动计算匹配过程所需的各种阈值；提出了一种大幅面遥感图像自动快速拼接方法，在下采样图像中提取特征和参数优化，把优化后的匹配对坐标直接传递到原始图像进行配准和拼接，并对常规的拼接缝消除方法进行了改进，提高了图像拼接质量，实现了遥感图像的全自动拼接，缩短了拼接时间，提升了图像拼接的自动化水平。为了解决常规的记录系统速率低与图像数据持续传输速率高之间的矛盾，在深入研究各种大容量数据记录系统优缺点的基础上，提出一种多通道软同步高速直写硬盘的数据记录方法，设计了自启动同步直写控制器，优化了SCSI硬盘的记录时序，有效地提高了硬盘的启动速度和数据持续记录速率，为实现每秒200兆字节高速大容量遥感图像数据的实时记录提供了技术支持。最后，对课题所研究的特征描述符构建方法、自动匹配策略、大幅面遥感图像拼接模型和多通道软同步高速数据记录系统进行了编程实现、设备制作和性能测试，实验结果显示：本文提出的基于扇型区域分割的特征描述符，在匹配性能与原SIFT算法相当的情况下，综合匹配时间得到了显著提高；自适应对比度阈值控制的方法有效增加了特征提取算法对遥感图像的适应能力，并且提取到的特征点分布均匀；基于改进的RANSAC自动匹配策略性能稳定，所用实验图像的平均查全率达到98.3%；大幅面遥感图像自动拼接方法得到的融合图像质量满足实际要求，拼接区平滑过渡，大幅度缩短了拼接处理时间。本文提出的多通道软同步高速数据记录技术，实现了每秒200兆字节高速大容量遥感图像数据的实时无损记录，比同类记录系统的速率提高了2倍以上，且体积小、使用方便，具有良好的可扩展性。更多还原

【Abstract】 Remote sensing image mosaic and recording technique is a key technique usedin the field of information-based reconnaissance equipment research anddevelopment to obtain high-resolution and large-scale battlefield images. It includesimage mosaic technique based on invariant feature and hard-disk direct writingrecording technique based on soft synchronization. However, the existing imagemosaic methods have such weak points as poor stability and adaptability, lowrecording speed, not highly automated operation, and very complicated structureand excessively high dimensions of feature descriptors. And none of the existingrecording techniques has a high enough continuous recording rate, which has adirect influence on further improvement of remote sensing reconnaissanceperformance. Therefore, this thesis intends to study fully automatic mosaic andreal-time continuous recording technique for images obtained using high resolutionimaging sensors, and to formulate a fully automatic mosaic technique for large scaleremote sensing images, and to find a high-speed recording system with a longcontinuous recording time suitable for reliable operation under asperityenvironments, thereby providing theoretical and technical bases for furtherimprovement of China’s remote reconnaissance system performance. This researchalso benefits space exploration, space remote sensing data processing, radarimaging data processing, radar imaging data processing and anti-SAR.This thesis starts with a summary presentation of the basic structure of aremote sensing image mosaic system, and the derivation of image transportmathematic model based on the basic principle of optical imaging, and a systematicexposition of key techniques, such as matching, transform, correction and fusion ofimages used for large scale image mosaic.Scale invariant feature transform (SIFT) is introduced though the in-depthstudy of conventional image feature extraction techniques for the extraction of largescale remote sensing image features. In light of the imperfect invariance andexcessively high dimensions of SIFT descriptors, a new way of constructing SIFTdescriptors based on sectorial segmentation is proposed to avoid the rotationaloperation during the construction of descriptor, and the number of dimensions is reduced from128to48while the invariance of descriptor is holden. In light of thepoor adaptability of SIFT descriptor with respect to an image with low contrast, aself-adaptive contrast threshold is proposed for the extraction of key points, therebyenhancing the adaptability of SIFT algorithm to a remote sensing image, i.e.expanding its scope of applicability.In light of the large numbers of SIFT features and large-scale remote sensingimages, three automatic matching strategies based on statistics, transform modeland improved RANSAC are proposed for the automatic calculation of thresholdsneeded for input images in the matching process. A fast automatic remote sensingimage mosaic process is proposed so that the optimization matching coordinate canbe transferred immediately after feature extraction in the sub-sampled images andparameter optimization to the original images for matching and mosaic, and theconventional mosaic seams are eliminated in the process as well. The new methodreduces the processing time and increases the automation level of remote sensingimage mosaic.In order to solve the conflict between the low recording speed of aconventional recording system and the high continuous transfer speed of image data,a hard-disk high-speed direct rating recording method based on multi-channel softsynchronization is proposed through a comparative study of different large capacitydata recording systems. The self-starting speed and the continuous recording speedof hard-disks are effectively increased by re-designing the self-startingsynchronization direct writing controller and optimizing the recording sequence ofSCSI hard-disks. A technical support is thus provided for the realization of real-timerecording of remote sensing image data at200MB/s.With programs developed and equipment produced, experiments are made withfeature descriptors, made in a new way automatic matching strategies, a large-scaleremote sensing image mosaic model, and the multi-channel soft synchronizationhigh-speed data recording system proposed in this thesis for performance evaluation.Experimental results indicate that the overall matching time has been improvedobviously by using the feature descriptors proposed while the matchingperformance is equivalent to that of original SIFT algorithm; the self-adaptivecontrast threshold control method has been used to effectively improve theadaptability of feature extraction algorithm to remote sensing images, and to make the distribution of feature points extracted even; the automatic matching strategybased on improved RANSAC has a stable performance, and the average reeall ratioof experimental images used goes up to98.3%; the quality of fused images realizedusing large scale remote sensing image mosaic method satisfied the actual needs,the transition between mosaic regions is smooth, and the stitching time is greatlyreduced. The multi-channel soft synchronization high-speed recording techniqueproposed in this thesis is used to realize the real-time nondestructive recording oflarge scale remote sensing images at200MB/s, more than double the speed ofsimilar recording systems, and it is small in size, easy to use, and has very goodexpandability as well.更多还原