一种改进的A-KAZE算法在图像配准中的应用

发布时间：2018-07-10 01:07

本文选题：A-KAZE + 非线性扩散滤波　；参考：《东南大学学报(自然科学版)》2017年04期

【摘要】：针对现有图像配准过程中难以保持图像的局部精度和边缘细节的问题,在A-KAZE算法的基础上提出了一种改进的图像特征提取算法AKAZE-ILDB.该算法首先利用非线性扩散滤波方程构造图像金字塔,采用快速显示扩散(FED)求得数值解,得到具有亚像素精度的图像特征点坐标;然后利用改进的LDB(ILDB)描述子构造具有尺度和旋转不变性的图像特征向量,对特征向量采用汉明距离进行KNN匹配;最后基于仿射变换模型计算空间映射参数矩阵来实现图像配准.实验结果表明:在保持相同图像特征匹配正确率的情况下,AKAZE-ILDB算法比A-KAZE算法平均配准时间缩短了300 ms;在配准精度方面,比A-KAZE算法提高了3.7%,比传统特征提取算法SURF匹配正确率提高了29%.
[Abstract]:In order to solve the problem that it is difficult to maintain the local accuracy and edge details in the present image registration process, an improved image feature extraction algorithm, AKAZE-ILDB, is proposed on the basis of A-KAZE algorithm. Firstly, the image pyramid is constructed by nonlinear diffusion filtering equation, and the numerical solution is obtained by fast display diffusion (Fed), and the coordinates of image feature points with sub-pixel accuracy are obtained. Then the improved LDB (ILDB) descriptor is used to construct the image feature vector with scale and rotation invariance, and the eigenvector is matched by hamming distance. Finally, the spatial mapping parameter matrix is calculated based on the affine transformation model to achieve image registration. The experimental results show that the average registration time of AKAZE-ILDB algorithm is 300 Ms shorter than that of A-KAZE algorithm under the condition of keeping the same image feature matching accuracy. Compared with A-KAZE algorithm, the algorithm improves 3.7 steps and 29% higher than the traditional feature extraction algorithm surf.
【作者单位】：东南大学计算机科学与工程学院;东南大学计算机网络和信息集成教育部重点实验室;东南大学蒙纳士大学苏州联合研究生院;
【基金】：国家高技术研究发展计划(863计划)资助项目(2015AA015904)
【分类号】：TP391.41

【相似文献】