一种精确图像待配准区域的快速配准算法

A Fast Registration Algorithm for the Region to be Registered

在众多图像配准算法中,尺度不变特征转换(Scale-invariant feature transform)由于其对旋转、尺度缩放、亮度变化保持不变性,且对仿射变换和噪声也有一定程度的鲁棒性,从而得到了大范围的应用。但该算法的缺点是运行时间过长。论文旨在减少不必要的检测区域,减少特征点检测和描述符构建所花费的时间。主要分为两部分:1)对待检测图像循环分块,计算各个子块信息熵(Entropy of information),当所有子块信息熵标准差达到固定阈值时停止分块,去掉子块集合中子块信息熵满足条件的部分。2)将图像二值化,计算二值图像连通区域,在第一步的基础上去掉所有连通区域的部分,得到最终待检测的图像区域。最后进行SIFT算法特征检测并生成PCA-SIFT描述子,使用最近邻与次近邻之比的算法完成图像配准。实验结果表明:论文算法在保证匹配精度达到90%以上的情况下,减少算法运行时间20%~30%,提高了图像配准的速度。

Among many image registration algorithms,scale invariant feature transform has been widely used because of its invariance to rotation,scale scaling,brightness change and robustness to affine transform and noise.But the disadvantage of this algorithm is that it runs too long.The purpose of this paper is to reduce the unnecessary detection area and the time of feature point detection and descriptor construction.It is mainly divided into two parts.Each sub block information entropy(entropy of information)is calculated when the standard deviation of all sub block information entropy reaches a fixed threshold,and the sub block information entropy satisfying the conditions is removed.The image is binarized,the connected region of binary image is calculated,and all the connected regions are removed based on the first step to get the final image region to be detected.Finally,the feature of SIFT algorithm is detected and PCA-SIFT descriptor is generated,and the ratio of nearest neighbor to next neighbor is used to complete image registration.The experimental results show that the algorithm reduces the running time by 20%~30%and improves the speed of image registration.