利用局部方向微分向量一致性的角点检测

在边缘轮廓提取的基础上,提出了一种利用局部方向微分向量一致性的角点检测算法以消除边缘噪声对角点检测产生的不利影响。该算法提取图像的边缘轮廓来降低算法计算量;利用各向异性高斯方向导数(ANDD)滤波器提取每个像素处的方向微分向量并进行幂次变换,以增强向量的各向异性;进而利用相邻像素的方向微分向量构建一致性测度。最后,对同一轮廓上的一致性测度进行均值归一化,得到最终角点测度。实验显示,提出算法的平均角点定位误差为1.52pixel,与对比算法接近;检测准确率分别比点到弦距离累积(CPDA)法、相对局部曲率(HeYung)法提高了58%和5.5%,与归一化残余面积(RA)算法相等,同时角点错检率比HeYung和RA少25.5%和21.6%。提出的算法能准确地检测出真实角点,并具有更小的错误检测率,更高的角点重复率,而且对边缘噪声十分鲁棒。

Corner detection via consistency of local directional differential vectors

On the basis of edge contour extraction, a new corner detection algorithm via the consistency of local directional differential vectors of pixels was proposed to eliminate the adverse effects of the edge noise on the corner detection. In the algorithm, the edge contour of an image was extracted to reduce the computation. Then, the Anisotropic Gaussian Directional Derivative( ANDD) filters were used to extract the directional differential vectors at each pixel on contours and the power transformation was employed to the vectors to enhance their anisotropy. Furthermore, the consistency was constructed from the directional differential vectors of adjacent pixels. Finally, the consistency was normalized by the average consistency on the same contour to produce the final corner measure. The experimental results show that the proposed algorithm achieves the average localization error of 1.52 pixel approximately to comparison algorithms. Also the detection accuracy is improved by 58% and 5.5% as compared with those of the Chord-to-point Distance Accumulation (CPDA) algorithm and relative local curvature (He&Yung ) algorithm and it is equal to that of the Residual Area (RA) algorithm. Meanwhile the false detection ratio is 25.5%, 21.6% lower than those of the He & Yung and RA algorithms, respectively. The proposed algorithm accurately detects the true corners in the image and holds a smaller false detection ratio and higher corner repeatability, and it is very robust to the edge noise.