一种快速检测图像角点特征的线搜索式方法

传统的图像角点特征检测方法在速度和准确性两方面难以兼顾. 针对该问题, 提出了一种角点特征检测的线搜索式方法. 该方法作用于一个以当前像素为中心核的圆掩模, 在该掩模内搜索通过核的所有直线, 如果存在一条直线不穿过核附近给定邻域以外的其他同值收缩核(Univalue segment assimilating nucleus, USAN)区域, 则当前像素点为角点. 论文论证了使用有限数目搜索线的可行性与必要性. 采用由粗及细的搜索策略, 动态设计搜索线的数目与搜索线上的检测点数目, 以提高检测速度. 提出了一种基于最大同值距离的新型非极大值抑制进行角点的精确定位, 并结合多种新型伪响应抑制措施, 有效地提高了算法的准确度. 实验结果表明该方法在准确性方面优于MIC、SUSAN和Harris等算法, 而且速度快, 仅稍慢于MIC算法, 具有优良的综合性能.

A New Algorithm for Fast Corner Detection Using Line Search Mechanism

In image corner detection, classical approaches can not meet the demand of high accuracy but low computational complexity. A new line search algorithm for image corner detection is proposed in this paper. The new method employs a circle mask centered at the pixel in question, called the ``nucleus' of the mask, to search all the straight lines that pass through the nucleus within the mask. Then, a corner point of an image is defined as the point (nucleus) where, among those straight lines, there is at least one straight line that does not cross the remaining part of the univalue segment assimilating nucleus (USAN) except a given neighborhood of the nucleus. This paper asserts the practicability and necessity of using finite search lines. To reduce the computational complexity, it employs dynamic search lines, finite points on each search line, and the coarse-to-fine search strategy as well. A new non-maximum suppression for accurate localization is proposed, and some new false response suppression measures are incorporated for good detection. Experiments show good performance of the new algorithm in both accuracy and speed.