计算机科学
散列函数
相似性(几何)
数据挖掘
二进制代码
语义相似性
嵌入
情态动词
语义鸿沟
人工智能
理论计算机科学
模式识别(心理学)
机器学习
二进制数
图像(数学)
图像检索
数学
化学
计算机安全
算术
高分子化学
作者
Fan Yang,Qiaoxi Zhang,Fengying Ma,Xiaoqing Ding,Yufeng Liu,Deyu Tong
标识
DOI:10.1016/j.ins.2023.119222
摘要
With its merits in query speed and memory footprint, hashing has elicited considerable monument in cross-media similarity retrieval applications. Many label-dependent supervised hashing methods have been proposed for similarity searching across modality boundaries. However, the current cross-modal hashing (CMH) works are subjected to severe information loss, and their performances may dramatically degrade because of the expensive costs of constructing affinity graphs, inadequate mining of label information, and disregarding label correlations. To facilitate these problems, we propose an Efficient Discrete Cross-Modal Hashing (EDCH) in which an asymmetric model is introduced, which not only conveys external semantic information via embedding high-order labels but also preserves internal modality attributes by introducing binary representations and common subspace. To fully use label semantic information, we integrate the semantic supervised intersection scheme and the category correlations embedding in a shallow framework. Moreover, we elaborately develop an efficient and effective discrete optimization strategy to learn binary representations and a novel mutual linear projection to strengthen the capability and effectiveness of hash functions. Comprehensive experiments are conducted on three representative datasets to evaluate our method. The results validated that our method achieved promising and competitive retrieval performance and surpasses several typical and cutting-edge approaches.
科研通智能强力驱动
Strongly Powered by AbleSci AI