具有近似仿射尺度不变特征的快速图像匹配  

Fast image matching algorithm with approximate affine and scale invariance

作　　者：岳娟[1,2] 高思莉 李范鸣[1] 蔡能斌 YUE Juan;GAO Si-li;LI Fan-ming;CAI Neng-bin(Key Laboratory of Intelligent Infrared Perception, Shanghai Institute of Technical Physics, Chinese Academy of Sciences , Shanghai 200083, China;Shanghai Key Laboratory of Crime Scene Evidence, Shanghai 200083, China)

机构地区：[1]中国科学院上海技术物理研究所智能红外感知重点实验室,上海200083 [2]上海市现场物证重点试验室,上海200083

出　　处：《光学精密工程》2020年第10期2349-2359,共11页Optics and Precision Engineering

基　　金：国家十三五预研项目资助(No.HJJ2019-0089/YYAA0089);上海市现场物证重点实验室开放课题资助项目(No.2017XCWZK08,No.2016XCWZK21)。

摘　　要：为了解决大仿射形变场景下,尺度不变特征变换(Scale Invariant Feature Transform, SIFT)算法的局限性以及仿射尺度不变特征变换(Affine-SIFT, ASIFT)算法运算量大的问题,提出了一种具有近似仿射尺度不变特征的快速图像匹配算法(Fast Approximate-Affine-SIFT, Fast-AASIFT)。该算法具有比ASIFT算法更清晰的物理意义,首先从逆仿射变换出发,对原图进行仿射形变纠正,估计出对应的正射图像;然后在正射图像上进行特征点提取及SIFT描述;最后进行SIFT优化匹配。实验结果表明:大仿射形变场景下,Fast-AASIFT算法依然能匹配到足够多的特征点,且峰值匹配误差<2.5 pixel,平均匹配误差<1.2 pixel,其抗仿射形变能力明显优于SIFT算法,与ASIFT算法相当;Fast-AASIFT算法耗时<0.3倍ASIFT,有效改善了ASIFT算法的耗时问题。可见,Fast-AASIFT算法既有效保证了算法抗仿射形变鲁棒性,又大幅提高了算法效率,对场景重构、场景识别等应用具有重要意义。To address the limitations of the scale invariant feature transform(SIFT) algorithm and reduce the computational burden of the Affine-SIFT(ASIFT) algorithm in scenes with large affine deformations, a fast image matching algorithm based on approximate-affine-SIFT(Fast-AASIFT) is proposed. Fast-AASIFT has a clearer physical meaning than the ASIFT algorithm. First, Fast-AASIFT recovers original images as rectified images by performing inverse affine transformations. Then, it performs feature point extraction and SIFT description on the rectified images. Finally, it performs SIFT optimization matching. The experimental results demonstrate that, in scenes with a large affine deformation, Fast-AASIFT can still match enough feature points, with a peak matching error of <2.5 pixels and an average matching error of <1.2 pixels. This proves that the anti-affine deformation ability of Fast-AASIFT is equivalent to that of the ASIFT algorithm, which is significantly better than that of the SIFT algorithm. Furthermore, the time consumed by Fast-AASIFT less than 30% of that consumed by the ASIFT algorithm;thus, it effectively addresses the time-consumption problem of the ASIFT algorithm. Obviously, Fast-AASIFT not only maintains good robustness against affine deformations but also greatly improves computational efficiency;consequently, it is of great value for applications such as scene reconstruction and recognition.

关 键 词：仿射不变性 图像匹配 仿射尺度不变特征变换 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]