摘要
为了提升包含2种场景类型的遥感图像的配准拼接质量,提出基于误差权重再分配的遥感图像配准拼接优化方法.使用尺度不变特征变换(SIFT)特征检测算子,提取2幅具有重叠区域的遥感图像的特征点,计算得到初始的单应性矩阵.针对遥感图像细节丰富,但在某些特定区域分布不均匀的特点,将图像按照网格分割成若干小子块,进行信息熵聚类.图像熵反映的是灰度分布的分散程度,较大的熵意味着更大的信息量和纹理细节.按照信息量分布,将图像分隔为2个大的图像区域,每一区域近似代表一种场景类型.以特征点匹配的残余误差为目标函数,对不同场景区域的特征点分配不同的优化权重,权重来源于各图像子块的信息熵,反映了图像各场景信息量的多少,从而改善拼接效果,使之符合人眼视觉要求.实验表明,采用该方法可以再分配特征点匹配残余误差,细节丰富区域的匹配残差降低14%,提升细节丰富区域的配准拼接质量,降低随机性,提高了配准过程的稳定度.
An optimization approach based on weighted value reassignment in RMSE was presented for remote sensing image registration and stitching in order to improve the registration and stitching quality of remote sensing image which consists of two scene types. Scale-invariant feature transform (SIFT) detector was used to extract feature points and obtain the homogeneous matrix between two images which have overlapped region. Remote sensing image has lots of details that may not evenly distribute. A method based on sub-block image entropy clustering to segment the image into large blocks was proposed. The image entropy reflected the dispersion degree of gray level distribution, and larger entropy value meant larger amount of information and more texture detail. Each segment represented a unique scene type and a weighted value would be assigned to feature points. In objective function of RMSE, the weighted value reflected the amount of texture information and contributed to the improvement of optimization result, which friendly matched the human vision. The experimental results showed that the registration residual error of feature points could be reassigned and was reduced by 14% in the region which had rich details. Then the registration and stitching quality and the stability of the registration process were improved, and the randomness was reduced.
作者
许越
徐之海
冯华君
李奇
陈跃庭
徐毅
赵洪波
XU Yue;XU Zhi-hai;FENG Hua-jun;LI Qi;CHEN Yue-ting;XU Yi;ZHAO Hong-bo(State Key Laboratory of Modern Optical Instrumentation, Zhejiang University, Hangzhou 310027, China;Shanghai Institute of Satellite Engineering, Shanghai 200240, China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2019年第1期107-114,共8页
Journal of Zhejiang University：Engineering Science
基金
国家重点研发计划资助项目(2016YFB0500803)
