文献类型：期刊文献

英文题名：A fast and precise spatial verification strategy for duplicate image retrieval

作者：Chen, Ming[1]; Zhang, Zhifeng[1]; Gao, Tieliang[2]; Duan, Li[3]; Zhang, Junpeng[4]

第一作者：Chen, Ming

通讯作者：Chen, Ming

机构：[1] Software Engineering College, Zhengzhou University of Light Industry, Zhengzhou, 450001, China; [2] School of Business, Xinxiang University, Xinxiang, 453000, China; [3] School of Computer and Information Technology, Beijing Jiaotong University, Beijing, 100044, China; [4] Zhumadian Power Supply Company, State Grid Henan Electric Power Company, Zhumadian, 463000, China

第一机构：Software Engineering College, Zhengzhou University of Light Industry, Zhengzhou, 450001, China

年份：2020

卷号：16

期号：9

起止页码：1393-1403

外文期刊名：International Journal of Performability Engineering

收录：EI(收录号：20204309379699);Scopus(收录号：2-s2.0-85092776032)

基金：This research was supported by the National Science Foundation of China (No. 61975187, 61902021) and Henan Science and Technology Research Project (No. 162102210214, 192102210294).

语种：英文

外文关键词：Image retrieval

摘要：Spatial verification for duplicate image retrieval is often time-consuming and not sensitive to similar images. To address this problem, we propose a fast and precise spatial verification strategy for duplicate image retrieval. The motivation of this strategy is to use angle and scale information to filter similar images. This is because the matched descriptors of similar images are not transformed according to consistent angles and log-scales. The angle differences and log-scale differences of matched descriptors can be projected to points in two-dimensional space. Intuitively, the two-dimensional point distribution of non-duplicate images is relatively discrete, and the two-dimensional point distribution of duplicate images is relatively concentrated. Therefore, this paper utilizes the inverse cloud algorithm to calculate the discrete degree of the two-dimensional point distribution to exclude the non-duplicate images that have large fluctuation distributions. Then, the new voting algorithm can be used to re-rank the images to improve the retrieval accuracy. The experimental results showed that, compared with traditional algorithms, the new strategy was able to effectively improve retrieval accuracy without adding extra storage overhead and computational overhead. ? 2020 Totem Publisher, Inc.