【作者】 曹世翔； 江洁； 张广军； 袁艳；

【Author】 CAO Shi-xiang,JIANG Jie,ZHANG Guang-jun,YUAN Yan (Key Laboratory of Precision Opto-mechatronics Technology,Ministry of Education.,School of Instrumental Science and Opto-electronics Engineering,BeiHang University,Beijing 100083,China)

【机构】 北京航空航天大学仪器科学与光电工程学院精密光机电一体化技术教育部重点实验室；

【摘要】 提出一种减少图像层数简化SIFT的图像配准方法。图像配准需要的正确匹配点对原则上只需要四对,另外误匹配点对透视投影矩阵参数的估计精度影响较正确匹配点大。从SIFT算法的热扩散方程出发,对图像金字塔进行比较,发现图像大部分的特征点都集中在第一层上。因此在RANSAC(一致稳健参数估计)确保透视投影参数求解下,去掉多余的Gauss和DOG(Difference Of Gauss)图像金字塔。实验数据表明此法能够减少误匹配点对且变换矩阵的计算精度和处理速度相对于传统SIFT均有提高。更多还原

【Abstract】 A simplified SIFT algorithm,which condensed the layers of classical SIFT,for image registration is proposed.Only four pairs of correct feature matches will constraint the registration in prince,besides,accuracy of model parameter evaluation between different points set is mainly affected by the percent of incorrect matching.It can be inferred from the heat diffusion equation that most of feature matches locate on the first layer.Thus,reduction of Gauss and DOG layers can eliminate incorrect match,meanwhile,with considerable improvement in perspective transform precision and runtime using RANSAC.Simulation demo proves its outperformance.更多还原