图像局部特征识别中的多目标分离  

Separate of Multi-objects in Image Recognition by Local Features

作　　者：吕冀[1] 汪渤[1] 高洪民[1] 周志强[1] 

机构地区：[1]北京理工大学信息科学技术学院自动控制系,北京100081

出　　处：《光子学报》2008年第8期1708-1712,共5页Acta Photonica Sinica

基　　金：国防预研项目(51405030104BQ0171)资助

摘　　要：研究了一种多目标识别算法，该算法用SUSAN角点形成SIFT特征点，采用阶梯图像金字塔结构实现尺度不变，为所有匹配点建立统一的超定线性方程组并对该方程组系数矩阵进行简化使其维数降低一半，得到增广矩阵．对增广矩阵进行列变换，依据坐标转换的特性可从中提取多目标的稳定正常点，实现了快速分离多目标的匹配点．结果表明，利用新算法得到的多目标识别结果能保证最小二乘法迭代运算快速收敛，且一次迭代就能得到精度较高的目标定位参量，根据SIFT标准的128维局部特征描述符判别匹配点，匹配点数量较SIFT算法多一倍，分离多目标速度较Hough变换快2～3倍．A multi-objects recognition algorithms which achieve rapid separation of all objects of the match points is proposed. Using SUSAN formed SIFT and ladder-image pyramid structure to achieve the same scale,new algorithm establish a unified set of over-determined equations of all match points. The,overdetermined linear equations coefficient matrix can be simplified to drop dimension, so simplify the structure of the augmented matrix and whose dimension is reduced to half. The inliers of all objects can be refined according to image transformation properties, and the inliers ensure robustness of the least-squares solution. The results shown that the multi-objects recognition algorithms the least squares iterative is convergence quickly,and the one step iteration will be able to get high accuracy orientation parameters. To discriminate inliers based on SIFT standard 128-dimensional local features,the number of match points is more than doubled SIFT algorithm. The multi-objects separated algorithms of augmented matrix can clearly separate multi-objects and is 2-3 times faster than Hough transform.

关 键 词：图像处理 目标识别 多目标分离 定位 异常点 增广矩阵 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]