改进的视角鲁棒KAZE特征匹配算法

Improved KAZE feature matching algorithm with viewpoint change robustness

针对KAZE特征匹配算法对视角变化敏感,在大视角场景下不能实现正确匹配的问题,提出了一种视角鲁棒的PKAZE（Perspective-KAZE）算法。该算法在原KAZE描述符的基础上,计算特征点邻域内的二阶梯度均值,形成新的扩展的80维描述符;然后利用透视变换模型对待匹配影像进行多视角模拟,在模拟影像上提取改进的KAZE描述符,再进行特征匹配。最后,选取5对含有最多正确匹配数量的影像上的匹配对作为初始结果,利用随机抽样一致算法对初始结果提纯。对多组图像进行了匹配实验,结果表明：与KAZE、尺度不变特征变换（SIFT）和加速鲁棒特征（SURF）算法相比,所提算法对视角变化具有更强的鲁棒性;与透视尺度不变特征（PSIFT）和仿射尺度不变特征（ASIFT）算法相比,本算法匹配正确率更高,分别为PSIFT的2-10倍,ASIFT的2-7倍。提出的算法对视角变化具有很好的鲁棒性,不仅对模拟影像的视角变化很稳健,而且适用于真实三维复杂场景拍摄的大视角影像,具有一定的实用价值。

The KAZE algorithm is sensitive to image view changes, so that it is not capable of the correct match between images with large view-point differences. This paper proposes a PKAZE （Perspective-KAZE） algorithm which is robust to the viewpoint changes. Firstly average second-order gradient values in the neighborhood region of key-points extracted by ordinary KAZE descriptors were calculated to extend the original KAZE feature descriptor to be a new 80-dimension one. Then, a perspective transform model was used to warp the image pairs to be matched in a series of different perspective angles. New KAZE feature descriptors were extracted on transformed image pairs and were matched later. Finally, five simulated image pairs with the most correct match numbers that are in top five simulated image pairs were selected as initial matching results, and the RandomSample Consensus （RANSAC） algorithm was used to remove false matching pairs in initial results. The matching experiments were performed on several image groups. The experimental results show that the proposed algorithm is more robust to the viewpoint changes as compared with the common KAZE, Scale Invariant Feature Transform（SIFT） and Speed Up Robust Feature （SURF） algorithms. The correct rate of proposed algorithm is 2-10 times that of the Perspective Scale Invariant Feature Transform（PSIFT） and 2-7 times that of the Affine Scale Invariant Feature Transform（ASIFT）. Furthermore, the proposed algorithm is not only robust to the viewpoint changes for simulated images, but also robust to the large view-point difference images in a real 3D complex scene. These results verify that the algorithm has a very good practical value.