基于Gabor多尺度空间的不变兴趣点检测

针对以往仿射不变兴趣点的特征尺度不能直接断定的问题, 提出一种基于Gabor多尺度空间的不变兴趣点检测算法。该算法主要包括三个步骤：应用Gabor滤波器组与图像卷积建立图像Gabor多尺度空间; 通过极大值准则检测兴趣点并直接断定特征尺度; 采用二阶矩矩阵描述兴趣点局部结构。实验结果表明, 相比较其他Hessian-Affine、MSER等算法, 该算法在图像模糊和JPEG压缩情况下可重复率和可匹配率均取得最好结果, 是一种能有效直接提取特征尺度的兴趣点检测算法。

Interest point detection based on Gabor multiscale-space

To solve the problem that the existing affine interest point detection algorithms cannot directly determine the characteristic scale for interest point, this paper proposed an invariant interest point detector based on Gabor multiscale-space. Firstly, this algorithm built the Gabor multiscale-space representation by smoothing the image with a series of Gabor filters. Secondly, it used the maxima criterion to detect the interest points and determine the characteristic scale. Finally, it described the local structure shape of interest point by the second moment matrix. The experimental data demonstrate that the proposed detector obtains the best performance under image blur and JPEG compression in terms of repeatability and matching score, compared to other detectors, such as Hessian-Affine and MSER et al, and then it is an effective interest point detection algorithm with determining the characteristic scale directly.