【作者】 冯晓伟；

【导师】 田裕鹏；

【作者基本信息】 南京航空航天大学 ， 测试计量技术及仪器， 2008， 硕士

【摘要】 图像配准是图像处理的基本任务之一,用于将不同时间、不同传感器、不同视角及不同拍摄条件下获取的两幅或多幅图像(主要是几何形式上)进行匹配。图像配准是多种图像处理及应用,如图像融合、图像拼接、变化检测等的基础,配准效果将直接影响到后续图像处理工作。在对图像配准方法进行分析的基础上,本文对基于形状内容描述子的图像配准方法和基于灰度差分不变量的图像配准方法进行了研究。本文提出了一种新的特征描述子―形状内容描述子(SCD)。其基本原理是根据曲线上的点到特征点的角度划分区域,得到一个N维的特征描述向量。构建的形状内容描述算子不依赖于特征点的全局特征,对图像的光照、位移、视角、噪声等有较强的适应能力。在将上述的特征描述算法应用于图像配准时,为克服这种图像配准技术只适用于图像间存在小角度旋转(约为0～5°)的不足,采用局部梯度方向直方图来确定旋转角度。灰度差分不变量(GDI)具有平移和旋转不变性,并且对噪声具有较强的鲁棒性。基于灰度差分不变量的图像配准算法首先利用灰度差分不变量构造特征描述子,然后通过计算各象素描述子之间的欧氏距离来决定是否匹配。实验结果表明,这种方法对噪声和配准误差都有一定的鲁棒性。本文最后介绍了图像配准在图像拼接和图像融合中的应用。将本文提出的两种图像配准算法应用在图像拼接上,实验证明该算法能够有效地拼接普通相机拍摄的照片,并具有较好的稳定性和较高精度。在图像融合中,采用透明融合技术对可见光和红外热图像进行融合。更多还原

【Abstract】 Image Registration is a fundamental task in image Processing used to (mainly in geometrical) match two or more images obtained at different time, from different sensors or from different viewpoints. Image Registration is the foundation of many image processing and applications such as image fusion, image stitching, change detection. And the effect of registration works on the following processing steps directly. In this paper, two image registration methods are presented: the registration based on the Shape Context Descriptor (SCD) and the registration based on the Gray-value Differential Invariants (GDI).A novel shape context descriptor is presented. For each feature, the shape context descriptor divides 360°central angle (the centre is the feature point) equally into N parts, then according the angle values between the feature point and the edge points, get a 1×N vector. This new technique doesn’t depend on global information. It is robust to noise, scale and so on. To overcome the drawback that the registration algorithm is only suitable to small rotation angle, the orientation difference between two images is calculated using a local gradient orientation histogram.Another registration algorithms based on the Gray-value Differential Invariants is introduced. The differential invariants are robust to noise, translation and rotation. Firstly the differential invariants are used to form local invariant feature, then the matching degree is gained via Euclidean distance. The experiment results show this method is robust to noise and registration error.Finally, we apply the two proposed algorithms in image stitching and image fusion. In image stitching, the results from experiments show the validity, good stability, and high precision of the two algorithms. In image fusion, transparency fusion technique is applied in the fusion of the visible and infrared light images.更多还原