Abstract: On the building image processing with Scale Invarint Feature Transform(SIFT) descriptor, there will be a large number of falsely matching points. Aiming at this problem, the color and global information is introduced to improve the performance of SIFT descriptor. It introduces l1l2l3 model which is robust against light change and build log-polar coordinates. For each key point, it builds circular neighborhood to cumulate the value of l1, l2, l3. Color invariant descriptor can be constructed. Global descriptor can be constructed with the same method. The Euclidian Distances of SIFT color invariant descriptor and global descriptor will be as similarity measurement. Experimental results indicate that the improved SIFT algorithm can reduce mismatch probability of building images and improve matching results greatly.

Key words: stereo matching, Scale Invarint Feature Transform(SIFT), global information, building image

摘要： 采用尺寸不变特征变换(SIFT)算法对建筑物图像进行匹配时会出现大量误匹配点。针对该问题，在SIFT彩色不变描述子中融入颜色信息和全局信息。引入对照明变化具有一定鲁棒性的l1l2l3模型建立对数极坐标，对于每一个特征点，在设定的圆邻域内累积l1值、l2值、l3值以构造彩色不变描述子，将特征点的最大曲率作为特征量以构建全局描述子，并计算SIFT描述子、彩色不变描述子和全局描述子的欧式距离作为相似性度量。实验结果表明，改进SIFT算法可以降低建筑物图像的误匹配率，提高匹配效果。

关键词: 立体匹配, 尺寸不变特征变换, 全局信息, 建筑物图像

CLC Number: 

• TP391