一种新的局部不变特征检测和描述算法  

A Novel Local Invariant Feature Detection and Description Algorithm

作　　者：杨恒[1] 王庆[1] 

机构地区：[1]西北工业大学计算机学院,西安710072

出　　处：《计算机学报》2010年第5期935-944,共10页Chinese Journal of Computers

基　　金：国家自然科学基金(60873085);国家"八六三"高技术研究发展计划项目基金(2007AA01Z314;2009AA01Z332)资助~~

摘　　要：局部不变特征已经被成功地用来解决计算机视觉领域诸多实际问题.文中提出一种新的局部不变特征检测和描述算法,提取出的特征能够对旋转、尺度缩放、光照等变化,甚至弱仿射变换保持不变.一般说来,局部特征的提取分为特征检测和描述两个关键步骤.在特征检测阶段,首先在每一层尺度图像上提取Harris角点,然后在以Harris角点为中心的固定大小的搜索窗内搜索三维尺度空间的极值点作为局部特征点的位置和特征尺度,最后为每个特征点计算主方向.文中的特征检测算法具有良好的可重复率性能.在特征描述阶段,建立了梯度的距离和方向直方图来描述局部特征,文中的特征描述子不但具有良好的匹配性能,而且维数更低,十分有利于提高图像特征的匹配速度.大量的图像匹配与图像检索实验结果验证了文中算法的有效性.Local invariant features have been successfully applied in many applications in computer vision. This paper proposes a novel local feature detection and description algorithm. The features are invariant to image rotation, scale and illumination changes, and even can be invariant to weak affine transformations. In general, the local feature extraction process can be divided into two key steps which are feature detection step and feature description step. In the detection step, firstly, the Harris corners are detected in every scale level image. Secondly, the local scale-space extrema is searched within a window which is center-localized on the multi-scale Harris corners. Finally, the predominant orientation is computed for each keypoint. The proposed feature detection algorithm has good repeatability performance. In the description step, a novel local descriptor is created based on the gradient distance and orientation histogram （GDOH）. GDOH not only has good matching performance, but also has low dimensionality, which results in much faster feature matching speed. Extensive experimental results have demonstrated the effectiveness and efficiency of the proposed algorithm.

关 键 词：局部特征 特征检测 特征描述子 不变性 图像匹配 

分 类 号：TP393.08[自动化与计算机技术—计算机应用技术]