一种鲁棒的二维图像形状检索方法

Robust Approach for 2D Shape-Based Image Retrieval

以往的基于极坐标变换的轮廓描述方法都是以形状质心作为参考点,使得其对噪声过于敏感,轮廓边界发生轻微的变化就可能导致形状匹配的失败.为此,文中提出一种用于描述和匹配二维图像形状的鲁棒性方法.该方法以广义霍夫变换参数映射的极值点作为待检索图像的参考点,以此来建立对应的极坐标变换轮廓描述曲线,从而通过计算描述曲线间所包围的累积误差面积来实现图像匹配.在3个图像库上的检索实验表明文中方法在具有平移、尺度、旋转不变性的同时还具有较好的鲁棒性,可有效对存在部分遮挡和局部缺损的图像进行检索.

The geometric centre is taken as the origin in the traditional centroid-radii based shape description method, which makes it sensitive to noise, and slight changes in the boundary cause errors in matching. A robust method for 2-D shape description and matching is presented to solve this problem. It uses a polar transformation of the contour points to get the shape descriptor, and the maximum of the generalized Hough transform （GHT） mapping array is taken as the reference point. The experimental results on 3 benchmark test sets show that the proposed approach is invariant to translation, rotation and scaling. Furthermore, it achieves high performance in the retrieval of partially occluded and defect images.