Abstract: Detected corner points using Harris-Laplace in multiple scales cause the problem that the same structure is detected in certain range of scales, which leads to finding many redundant points. Those points increase the complexity of computing in the later corner point description and matching procedure. Meanwhile, the difference of those points in scales and locations also causes mismatch. An improved Harris-Laplace method is proposed, which selects one characteristic point to represent the same local structure. With the improved Harris-Laplace method and SIFT descriptor, through setting the threshold of maximum and second maximum of distance, the auto image registration is realized. The results of many experiments compared with the original method indicate that the improved method not only has better invariant performance in image rotation transform illumination change and scale transform, but also can obtain stable matching pairs. Except that, for getting rid of many redundant points in detecting procedure, the consumed time of image registration and mismatch probability are also reduced.

Key words: Harris-Laplace, redundant points, Scale-Invariant Feature Transform（SIFT） descriptor, image registration

摘要： 利用Harris-Laplace算法对一幅图像进行多尺度特征点检测时，图像的局部结构在一定的尺度范围内被多次检测到，从而产生冗余点。冗余点不但增加了后续配准的计算量，同时由于这些表示同一局部结构的冗余点在位置和尺度上的差异降低特征匹配精度导致误匹配。通过对表示局部结构的特征点进行选择，提出了Harris-Laplace的改进算法。利用改进Harris-Laplace算法结合SIFT描述子，通过设定最小距离与次最小距离的阈值实现了图像的自动匹配，与原来算法作了大量的对比实验。实验结果表明，该算法不仅具有更好的旋转、光照和尺度不变性还具有获得稳定数量的匹配点的特性。同时，由于该算法相对于原算法在特征检测阶段减少了大量的冗余点，所以提高了图像配准的速度并降低了误匹配。

关键词: Harris-Laplace, 冗余点, 尺度不变特征变换（SIFT）描述子, 图像配准