KAZE特征检测与描述算法在图像匹配方面具有较好的性能。然而, KAZE算法中Perona-Malik(P-M)模型的解不具有唯一性, 而且图像弱边缘在尺度空间中易被平滑。为此, 提出一种改进的KAZE特征检测描述算法(CKAZE)。首先, 基于KAZE原理与能量泛函构建自适应扩散滤波函数; 然后, 研究解的唯一性及图像滤波过程中的边缘保持能力; 最后, 提出CKAZE算法, 利用Mikolajczyk标准数据库图像进行特征匹配实验, 对其性能进行验证。结果表明, 对高斯模糊、光照、旋转缩放、视觉变换而言, CKAZE算法的特征匹配正确率分别较KAZE算法高4.555%、2.138%、0.656%、1.981%, 特征检测和描述的精度提高。

For image matching, the KAZE feature detection and description algorithm has demonstrated a number of advantages. However, the solution of Perona-Malik (P-M) model adopted by KAZE is not unique, and the weak edges of image are prone to be smoothed in scale spaces by nonlinear diffusion filter function when the feature points are detected. To overcome these problems, an improved KAZE feature detection and description algorithm for image matching (CKAZE) is proposed. Firstly, an adaptive diffusion filter is built based on the principle of KAZE and energy functional. Then, the solution uniqueness and the edge preserving capacity of the proposed adaptive diffusion filter function are studied during filtering process. Finally, the CKAZE is constructed and its performance is validated through image matching experiments on Mikolajczyk benchmark image dataset. The results demonstrate that the correct rates of feature matching through CKAZE is 4.555%, 2.138%, 0.656% and 1.981% higher, respectively, than those by KAZE for Gauss blurring, illumination, rotation zoom and visual transformation, which indicate that the accuracy of feature detection and description is improved by CKAZE.