| 引用本文: | 郑奇斌,刁兴春,王彦臻,等.跨模态检索中的相似性漂移问题.[J].国防科技大学学报,2021,43(5):99-106.[点击复制] |
| ZHENG Qibin,DIAO Xingchun,WANG Yanzhen,et al.Similarity drifting problem in cross-modal retrieval[J].Journal of National University of Defense Technology,2021,43(5):99-106[点击复制] |
|
| |
|
|
| 本文已被：浏览 7161次 下载 4729次 |
| 跨模态检索中的相似性漂移问题 |
| 郑奇斌1,2,刁兴春3,4,王彦臻3,4,曹建军5,刘艺3,4,秦伟3,4 |
|
(1. 陆军工程大学 指挥控制工程学院, 江苏 南京 210007;2. 军事科学院, 北京 100089;3. 军事科学院 国防科技创新研究院, 北京 100071;4. 天津(滨海)人工智能创新中心, 天津 300450;5. 国防科技大学 第六十三研究所, 江苏 南京 210007)
|
| 摘要: |
| 为了降低“相似性漂移”问题的影响,提出一种基于“邻域传播”的匹配策略,将待查询项的模态内近邻映射到目标空间中,并将它们在目标空间中的最近邻作为查询项的跨模态近邻。基于邻域传播的匹配策略在不改变跨模态映射函数的条件下,可以有效地降低“相似性漂移”带来的误匹配现象。理论和实验分析证明,跨模态映射函数的“相似性漂移”问题广泛存在,而基于“邻域传播”的匹配策略可以有效降低其影响,提高匹配的准确率。 |
| 关键词: 跨模态检索 相似性漂移 邻域传播 深度神经网络 |
| DOI：10.11887/j.cn.202105012 |
| 投稿日期：2020-02-25 |
| 基金项目:国家自然科学基金资助项目(91648204,61532007,61371196) |
|
| Similarity drifting problem in cross-modal retrieval |
| ZHENG Qibin1,2, DIAO Xingchun3,4, WANG Yanzhen3,4, CAO Jianjun5, LIU Yi3,4, QIN Wei3,4 |
|
(1. Command and Control Engineering College, Army Engineering University, Nanjing 210007, China;2. Academy of Military Sciences, Beijing 100089, China;3. National Innovation Institute of Defense Technology, Academy of Military Sciences, Beijing 100071, China;4. Tianjin Artificial Intelligence Innovation Center, Tianjin 300450, China;5. The Sixty-third Research Institute, National University of Defense Technology, Nanjing 210007, China)
|
| Abstract: |
| In order to reduce the impact of the “similarity drift” problem, a matching strategy based on “neighborhood propagation” was proposed, which maps the intra-modal neighbours of the query items onto the target space, and takes their nearest neighbours in the target space as cross-modal neighbours of the query term. Experiments on real data sets prove that the similarity drifting problem exists widely, and the proposed matching strategy can effectively reduce its impact and improve the accuracy of matching. |
| Keywords: cross-modal retrieval similarity drifting neighborhood propagation deep neural networks |
|
|
|
|
|
