College of Artificial Intelligence, North China University of Technology
Tangshan immersive virtual environment 3D simulation foundation innovation team(No.18130221A)
针对图像匹配中AKAZE(accelerated-KAZE)算法匹配精度较低以及计算复杂等问题,提出了一种基于高斯滤波和AKAZE-LATCH算法相融合的图像匹配算法。首先,对输入图像进行高斯滤波预处理,去除高斯噪声等连续性噪声,并且保留图像的边缘信息。然后通过LATCH(Learned Arrangements of Three Patch Codes)算法为AKAZE算法构建高效的二进制描述子,再通过KNN(K Nearest Neighbors)算法得到对应的匹配对。最后结合USAC(UniversalRANSAC)去除误匹配对方法进行再次筛选,得到最终的匹配结果。经实验对比,本文算法相较于AKAZE算法匹配精度更高,具有良好的鲁棒性和可靠性,可用于多数复杂场景下的图像匹配。
To address the problems of low accuracy and complex computation of AKAZE (accelerated-KAZE) algorithm in image matching, an image matching algorithm based on the fusion of Gaussian filtering and AKAZE-LATCH algorithm was proposed. Firstly, the input image is preprocessed by Gaussian filtering to remove continuous noise such as Gaussian noise and retain the edge information of the image. Then efficient binary descriptors were constructed for AKAZE using LATCH (Learned Arrangements of Three Patch Codes) algorithm, and corresponding matching pairs were obtained using KNN (K Nearest Neighbors) algorithm. Finally, the method was screened again with USAC(UniversalRANSAC) to remove false matches, and the final matching result was obtained. Experimental comparison shows that compared with AKAZE algorithm, the proposed algorithm has higher matching accuracy, good robustness and reliability, and can be used for image matching in most complex scenes.