图像制导中的SIFT快速算法  

Fast SIFT algorithm for autonomous image guidance

作　　者：吕冀[1] 高洪民[1] 汪渤[1] 周志强[1] 

机构地区：[1]北京理工大学信息科学技术学院,北京100081

出　　处：《系统工程与电子技术》2009年第5期1147-1151,共5页Systems Engineering and Electronics

基　　金：国防预研基金资助课题(51405030104BQ0171)

摘　　要：图像制导中,导弹根据预存的地形图自动捕捉攻击目标。通过图像匹配识别目标,尺度不变特征变换(the scale invariant feature transform,SIFT)算法具有优异的性能,但计算量较大。根据地形图像的特点提出了一种SIFT改进算法Zoser SIFT,Zoser SIFT直接对图像进行逐级递减采样,形成阶梯(Zoser)金字塔图层,由24邻域极值点形成特征点。新算法不进行高斯变换,具有图层少、特征点数量适中等优点,大幅度减少了计算量。同时新算法占用内存空间少,浮点运算少,适合在实时DSP系统中应用。新算法虽然抗噪声能力有所下降,但对光照改变、尺度变化、变形和遮挡等仍有很好的鲁棒性,在航拍地形图上进行识别时性能稳定。At autonomous image guidance systems, the missile recognizes targets according to the saved physiographic pictures, and SIFT is outstanding in image matching but has relatively computational load. The novel algorithm builds the step pyramid （the zoser pyramid） representation by successively reducing the image size with combined interpolation and does repeatedly not convolve the initial image with Gaussian at each scale of scale space. Each sample point makes a comparison with its 24 neighbors to detect the local maximum and minimum for each scale. It can exactly recognize objects in photo maps even though addition of the noise, affine distortion, and change in illumination. The algo rithrn processes more quickly and needs less memory units at the real-time on-board DSP system.

关 键 词：图像处理 目标识别 图像匹配 尺度不变特征变换算法 金字塔图层 特征点 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]