通信地址-西北大学智能信息处理试验室
PAGE \* MERGEFORMAT 2 光学 精密工程 第 卷
PAGE \* MERGEFORMAT 2
第 期 PAGE \* MERGEFORMAT 3冯筠,等:
PAGE \* MERGEFORMAT 3
收稿日期:xxxx-xx-xx;修订日期:xxxx-xx-xx
基金项目:国家自然科学青年基金项目(No
Vol. x No.x xxx. xxxx第 x 卷 第x 期xxxx年 x月
Vol. x No.x xxx. xxxx
第 x 卷 第x 期
xxxx年 x月
Optics and Precision Engineering
文章编号
一种利用学习不变特征变换的
兵马俑图像分区匹配方法
冯筠,延瑜瑜,赵妍,肖方,刘晓宁*
(西北大学 信息科学与技术学院,陕西西安 710127)
摘 要:针对兵马俑图像三维重建时特征误匹配率较高、效率较低的问题,提出了一种基于两视图兵马俑图像的特征分区匹配方案。该方案在兵马俑图像上,首先利用学习不变特征变换(Learned Invariant Feature Transform, LIFT)方法提取整幅兵马俑图像的特征,接着利用提出的基于先验知识的特征点分布曲线分割方法确定兵马俑头部分割线位置,根据头部分割线将提取的特征分为头部特征和躯干特征。最终采用欧式距离进行分区特征匹配,对于匹配结果集合使用随机抽样一致性算法(Random Sample Consensus, RANSAC)滤除误匹配点集。实验结果表明,在兵马俑图像特征提取及匹配中该方案的正确匹配率可以达到98%,相比于SIFT和SURF方法正确匹配率提高了20%左右,特征点的可重复率提高了10%左右,同时将RANSAC的迭代时间降低了50%,而且在尺度、光照、角度发生变换时具有较好的鲁棒性。因此本文提出的方案能够很好地实现特征点的正确匹配,在兵马俑的三维重建中具有很高的使用价值。
关键词:兵马俑;特征提取;特征匹配;LIFT;RANSAC
中图分类号:TP391 文献标识码:A
A Terracotta Image Partition Matching Method Based on Learned Invariant Feature Transform
FENG Jun,YAN Yu-yu,ZHAO Yan, XIAO Fang,LIU Xiao-ning*
(College of Information Science and Technology ,Northwest University,Xi'an 710127 ,China)
Abstract: This paper presents a novel feature partition matching scheme for two-view Terracotta Warriors images to alleviate the problem of higher false matching rate and lower efficiency for the Terracotta Warriors images feature matching in 3D reconstruction. The new scheme is that: First, the features of the whole Terracotta Warriors image are extracted by using the Learned Invariant Feature Transform (LIFT) method. Second, the position of the dividing line on the head of Terracotta Warriors is determined by applying the proposed prior knowledge-based feature point distribution curve, and then the extracted features are divided into the head features and the torso features based on the dividing line of the head. Third, the Euclidean distance is used to ca