【作者】 刘倩；

【导师】 梁毅；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2022， 硕士

【摘要】 图像处理技术是当前信息化时代的热点话题,多模态图像的涌现带来图像信息丰富度的扩增,如何有效利用异源图像获得更加全面而精确的地物目标信息,成为一项亟需解决的问题。异源图像配准融合技术可对包含同一场景信息的多模图像进行空间对齐与各层级信息整合,进而获得特征增强、内容丰富的融合图像。各传感器的工作模式各有千秋,其中可见光成像系统发展最为成熟,所获得的图像具有丰富的色彩信息和清晰的纹理背景,但是该成像系统易受天气、光线以及遮挡物的影响,使用场景受限严重。红外探测系统和SAR利用其自身的全天时、全天候优势,可以弥补光学成像系统的不足,获取目标对应的互补信息,但是红外图像和SAR图像的可读性较差,缺乏地物目标对应的光谱信息。为了实现多传感器的优势互补,提高融合图像的语义表征能力,本文主要开展以下工作:1.对多传感器下的地物特性和表征信息做了详细的对比剖析,分析影响异源图像配准性能的主要缘由。阐述图像配准技术和图像融合技术的原理框架,分析研究不同类别的图像配准融合方法以及各项评价指标的物理意义和数学定义,为后续开展异源图像配准融合技术夯实基础。2.针对异源图像之间存在的辐射差异、几何差异,本文利用异源图像频域相位特征的相似性和稳定性,提出基于相位一致性的异源图像配准方法,克服传统图像配准算法在复杂应用场景中易失配的问题,获得精度高、适用性广且实时性强的配准结果。首先,为了降低细小杂乱的噪声边缘对后续特征检测的影响,开展SAR图像滤噪算法的研究,通过实验对比以及性能分析,选取PPB滤波算法实现SAR图像的斑噪抑制。其次,本文提出基于相位一致性叠加矩的特征分块检测策略,使得最终的配准点对尽可能分布均匀,获得兼容性更强的几何变换参数。随后,为了适应基准图像和实时图像之间存在旋转缩放差异的配准场景,提出基于局部频域特征的异源图像配准方法,在该方法中采用相位一致性信息设计特征点主方向并构造混和描述子进行特征描述。最后,提出一种基于结构相似性的异源图像快速模板配准方法,在降低时间成本的同时提升配准精度。通过与当前该领域比较新颖且效果显著的配准算法进行对比,验证本文所提算法的优势之处。3.针对传统图像融合方法出现特征缺失、易受噪声干扰且灰度过渡不平滑等问题,本文提出一种基于多方向梯度滤波的异源图像融合方法。设计一种多方向边缘保持器,并结合拉普拉斯金字塔的多尺度优势,对源图像进行两阶段分解。在此基础上设计不同层级对应的融合规则,最大化信息利用率,最小化噪声干扰带来的影响,获得目标显著、结构明晰且光谱丰富的融合结果。通过与当前该领域经典的算法和融合策略进行对比,验证本文所提算法的优势之处。更多还原

【Abstract】 Image processing technology is a hot topic in the current information age.As an important branch of image processing,image registration and fusion technology has been widely used in remote sensing mapping,disaster monitoring,security inspection,and military guidance.With the advancement of science and technology,sensor imaging technology has become more and more mature,and the emergence of multimodal images has brought about an increase in the richness of image information,how to effectively use heterologous images to obtain more comprehensive and accurate object target information has become an urgent problem to be solved.Heterologous image registration and fusion technology achieves spatial alignment and information integration by performing distortion correction on multisensor images of different phases,different spectral bands and different spatial positions,and then obtains fusion images with enhanced features and rich content.Mechanisms of different sensors have their own strengths and weaknesses.Among them,the visible light imaging system is the most developed,and the obtained images have rich color information and clear texture backgrounds.However,the imaging system is easily affected by weather,light and obstructions,and does not have the advantage of all-day and all-weather.Infrared sensors and synthetic aperture radar(SAR)can break through the limitations of optical imaging systems and obtain target information that is difficult for optical sensors to obtain.However,the readability of infrared images and SAR images is poor,and the spectral information corresponding to ground objects is lacking.In order to complement the advantages of multi-sensor images and improve the information expression ability of fused images,the thesis conducts research on heterologous image registration and fusion technology.The main work is as follows :1.Detailed comparison and analysis of the ground object characteristics and characterization information under multi-sensors is carried out,and the main factors affecting the registration accuracy of heterologous images are analyzed.The thesis expounds the principle framework of image registration technology and image fusion technology,analyzes and studies different types of image registration and fusion methods,and the physical meaning and mathematical definition of each evaluation index,laying a solid foundation for the subsequent development of heterogeneous image registration and fusion technology.2.Aiming at the radiation differences and geometric differences between heterologous images,the thesis proposes a heterologous image registration method based on phase consistency,which utilizes stable image frequency domain features to obtain high-precision heterologous image registration results.Since the inherent speckle noise of SAR images will affect the registration results,the thesis first studies the filtering algorithm of SAR images,and selects PPB filtering algorithm to denoise SAR images through experimental comparison and performance analysis.Secondly,due to the existence of local distortion in SAR images,the thesis proposes a feature block detection strategy based on phaseconsistent superposition moments,so that the final registered point pairs can be distributed as evenly as possible,and more adaptable geometric transformation parameters can be obtained.Then,in order to adapt to the registration scene where there is a difference in rotation and scaling between the reference image and the real-time image,a heterologous image registration method based on local frequency domain features is proposed,the main directions of feature points are designed based on phase consistency information,and a mixed descriptor is constructed for feature representation.Finally,a fast template registration method for heterologous images based on structural similarity is proposed,which can improve the registration accuracy while reducing the time cost.The feasibility and effectiveness of the algorithm proposed in the thesis are verified by comparing it with the relatively novel and effective registration algorithms in the current field.3.Aiming at the problems that traditional image fusion methods are prone to feature missing,uneven grayscale transition and noise interference,the thesis proposes a heterogeneous image fusion method based on multi-directional gradient filtering.Design a multi-directional edge-preserving filter and combine the multi-scale advantages of the Laplacian pyramid to perform a two-stage decomposition of the source images.On this basis,the fusion rules corresponding to different levels are designed to maximize information utilization,minimize the impact of noise interference,and obtain fusion results with significant targets,clear structures and rich spectra.Finally,several sets of comparative experiments are designed to verify the advantages of the proposed algorithm from different perspectives.更多还原