一种适用于图像拼接的DSIFT算法研究

Research on a DSIFT Algorithm Applicable to Image Mosaicking

针对SIFT算法在分辨率很低的模糊边缘平滑图像中提取的特征点数量过少,而且没有考虑特征点的分布情况、计算开销较大的问题,提出了一种离散尺度不变特征提取DSIFT(Discrete SIFT)算法。该算法在空间极值检测阶段引入一个滑动窗口,在窗口内对极值点的检测进行非极大值抑制,使得特征点的分布相对均匀,运算速度更快,并且保持了尺度、旋转、仿射等不变性。在特征提取前添加了降采样操作,在计算单应矩阵前添加位置信息还原的步骤,在查找匹配点的过程中引入K-D树,以及在特征点的筛选和单应矩阵的估计上采用RANSAC算法,都降低了图像配准各个阶段的时间开销。最后,通过实验验证,DSFIT算法相对SIFT算法具有更加均匀的特征点分布,保持了较高的鲁棒性,同时,在保证一定图像拼接质量的前提下极大地降低了图像配准各个阶段的时间开销。

In view of the fact that the SIFT algorithm extracts the feature points too little and ignores their distribution, and also has the problem of high computation cost, a discrete-scale- invariant feature extraction algorithm （discrete SIFT or DSIFT in short） is proposed. This algorithm introduces a sliding window on the space extreme test phase of the SIFT algorithm, implements non-maximum suppression for the extreme points detection inside the window, so that the feature points are distributed relatively even. In addition, this algorithm makes the calculation faster, and at the same time maintains the scale, rotation, and affinity invariant. In order to reduce the time overhead of the image registration in various stages, it adds a desampling operation before feature extraction and the operation of location information reversion before calculating the homographic matrix. And it also introduces K-Dimensional Tree in the process of searching matching points, and adopts RANSAC algorithm in the shifting of the feature points and estimation of homographic matrix. Finally, through experiment verification, it is found that the DSIFT algorithm has more uniform distribution of feature points than SIFT algorithm, and with high robustness. At the same time, on the premise of guaranteeing the quality of image mosaicking, the time overhead in various stages of image registration is greatly reduced.